ComfyUI Workflows

iGEN ONE (ZImage, Ernie, FLUX1/2, Qwen, HiDream, Nunchaku)

iGEN ONE — Workflow Guide579+ nodes · 40 groups · 13 component subgraphs · 2 pipeline rows161 unique node types — 71% Eclipse nodesBuilt with ComfyUI_Eclipse custom nodesWhat Is This?iGEN ONE is a modular, all-in-one image generation and post-processing pipeline for ComfyUI. It supports a wide range of diffusion models — Flux, Stable Diffusion, HiDream, and more — and covers everything from initial image generation through face detailing, upscaling, and watermarking in a single workflow.The key design principle is modularity: every feature lives in its own group that can be independently enabled or disabled by simply muting or bypassing it. You never need to reconnect anything — the pipeline automatically adapts to whatever groups are active.How It Works — The BasicsLayoutThe workflow is arranged in two horizontal rows that you read left to right:Row 1 (27 groups) — Everything needed to generate an image: image inputs, prompts, model loading, and renderingRow 2 (13 groups) — Everything that happens after generation: refining, face swap, detailing, upscaling, watermarks, and savingBetween the two rows sits a routing banner — a strip of utility nodes that resolves shared resources (model, VAE, CLIP, prompts, image dimensions) so that every Row 2 group can find what it needs automatically.Toggling Features On and OffEach group has a Fast Mode Switcher panel — a small control panel that lets you mute or bypass individual sub-features within the group. Think of it like a row of toggle switches for that section's optional capabilities. To disable an entire group, you mute/bypass the group itself in the ComfyUI canvas.To mute/activate an entire group: right-click the group header → "Set Group Nodes to Never" (mute all) or "Set Group Nodes to Always" (activate all).When a group is muted/bypassed, downstream groups automatically skip it and pick up from the last active group. This works because of a priority-based fallback system: each group tries a list of possible input sources in order and uses the first one that's actually active.You can enable any combination of groups and the pipeline will always find the right data path. There is no need to manually reconnect anything.Data RoutingInstead of visible noodle connections between groups, iGEN ONE uses Set/Get nodes — named value channels that work like wireless connections. A SetNode in one group publishes a value (like "ref_image" or "model_init"), and a GetNode in another group retrieves it by name. This keeps the visual layout clean and makes it easy to rearrange groups.Row 1 — Generation PipelineImage Sources (Groups 1–3 + 5–9)The workflow offers four ways to get a starting image. Only one should be active at a time — the pipeline automatically picks whichever source is enabled. A loaded image can serve two purposes: as a visual reference for img2img generation, or simply as input for the Image to Prompt group (group 8) to generate a text description — you don't have to use it for img2img.You can also load an image and skip the Initial Render entirely — disable the Initial Render switch, and the loaded image goes straight to Row 2 for detailing, upscaling, face swap, or any other post-processing. This lets you bring in images from anywhere (other workflows, other tools, photographs) and run them through the full post-processing pipeline.1. Image LoadLoad a single image from disk. This is the simplest option — pick an image and go. It also extracts any embedded generation metadata (model name, prompt, sampler, seed) from the image, and can optionally override the workflow's settings with those extracted values. This is useful for "remix" workflows where you want to re-generate with the same settings that produced the original.2. Image Load from FolderBatch processing mode. Loads images one by one from a folder, with controls for sorting (by name or date) and subfolder traversal. Like Image Load, it can extract and apply metadata from each image. Great for re-processing an entire folder of images through the pipeline.Set the index to -4 for shuffle mode (random order, no repeats). The optional seed_input slot controls when special modes advance — connect a seed and keep it the same value to freeze the image selection while you tweak other settings. Change the seed value to advance to the next image.3. Input Video FrameExtracts a single frame from a video file, with a configurable frame skip offset. Useful when you want to use a video still as your starting image.4. Text-Only GenerationWhen none of the above are active, the workflow generates purely from text prompts using an empty latent. This is the default "txt2img" mode. Even when an image source is active, it only becomes img2img if you also enable one of the i2i sub-features in the Initial Render group (i2i Denoise, Flux Preproc, DiffSynth Qwen, etc.) — otherwise the loaded image is only used for reference purposes like Image to Prompt.After selecting a source, the image can pass through several optional processing steps:5. Remove BackgroundRemoves the background using BiRefNet, isolating the subject on transparency. Useful when the background would interfere with generation or when you want to focus on the subject only.6. Image Crop — AutoAutomatic subject-aware cropping. Uses SegmentAnything (SAM) to detect the main subject, centers the crop on it, and resizes to your target dimensions. Best for single-subject images where you want tight framing.7. Image Crop — CustomManual bounding-box cropping with pixel-level controls. For when auto-crop doesn't frame things the way you want.8. Preview Cropped ImageA preview checkpoint with a Stop node. Enable this to see the crop result and halt execution before proceeding — useful for verifying your crop settings.9. Resize ImageSimple resize to specific dimensions. Used when your input image doesn't match the target generation size.The image source chain has a built-in priority system: it checks from the last processing step backward (resize → crop_preview → crop_custom → crop_auto → rembg → video → folder → load) and uses the first active result. So you can stack processing steps and the last one wins.Prompt Construction (Groups 8, 18–23)There are multiple ways to build your prompt, and they can be combined:8. Image to PromptUses an AI vision model (Qwen 9B, Q4_K_M quantization) to analyze your reference image and generate a text description. Runs via Eclipse's Smart LM Loader with a "Detailed Description" task. The result feeds into the prompt assembly group as one of the possible prompt sources.18. RaffleRandom prompt generation from a curated tag system. Raffle builds prompts by randomly selecting tags from categories (subject, pose, clothing, etc.) with seed-controlled reproducibility. Includes a negative output filter for excluding unwanted content.19. Read Prompt from FilesReads prompts from external text files (one prompt per line). Uses index-based indexing to select which prompt to use. Good for working through a prepared list of prompts in sequence.Like Image Load from Folder, set the index to -4 for shuffle mode. Connect a seed to the seed_input slot and keep it fixed to freeze the prompt selection while tweaking other settings — change the seed value to advance to the next prompt.20. PromptThe central prompt assembly hub. This is where all prompt sources come together into the final positive and negative prompts.What's inside:Wildcard Processor — Template-based prompting with __wildcard__ placeholders for varietySmart Prompt v2 (Subject) — A structured subject builder with dropdowns for gender, age, hair, clothing, etc.Smart Prompt v2 (Settings) — Environment builder with dropdowns for location, time of day, weather, etc.Join nodes — Combines all active prompt inputs (from Image-to-Prompt, Raffle, file reader, manual text)String DeDuplicate — Automatically removes duplicate tags or phrases from the combined promptPrefix / Suffix — Optional quality tags (like "masterpiece, 8K") added before or after your promptNegative Prompt — A multiline text field for your negative promptEach prompt source has its own Mode Bridge toggle, so you can enable any combination: just the manual prompt, manual + raffle, image-to-prompt + files, or any other mix.The Prompt group has three sub-feature toggles that control other groups: enabling Image to Prompt activates the Image to Prompt group (you still need to manually enable an image source group like Image Load — follow the arrow from Image to Prompt back to find it), Raffle activates the Raffle group, and Read from Files activates the Read Prompt from Files group.21. Prompt StylerWraps your positive prompt in a style template. Uses Eclipse's Prompt Styler node to apply a consistent style (like "photo-hdr") to the prompt text.22. Prompt EditAI-powered prompt rewriting. Uses the same Qwen 9B model but with a "Rewrite Style" task — it takes your prompt and creatively rewrites it while preserving the core meaning. Good for generating variations or improving prompt quality.23. Save PromptsSaves the final combined prompt to a text file. Can append to an existing file, letting you build a collection of prompts over time.Model Loading & Enhancement (Groups 9–17)9. Folder / SizeThe configuration hub for the workflow. Sets:Output folder structure (with date-based subfolders)Image dimensions — default is 896×1152 (3:4 aspect ratio, good for portrait)Batch size — how many images to generate per runLatent type — SD3/Flux/Wan/HunyuanVideoVRAM purge behavior10. Model LoaderLoads the main checkpoint using Eclipse's Smart Model Loader. The default configuration loads Flux Kreamania fp16 in UNet mode with fp8_e4m3fn weight quantization and flash-attention2. Uses external CLIP models (ViT-L-14 + t5xxl_fp8) and an external VAE (flux_vae). The Smart Model Loader handles all the complexity of model configuration in one node. The usual main model is darkBeast Blitz8.The Smart Model Loader has a built-in template system — you can save your entire loader configuration (model, CLIP, VAE, sampler settings, etc.) as a named template and restore it later with one click. The workflow ships with pre-built templates, but those reference specific checkpoints you may not have. You can either download the matching model or load a shipped template and swap in your own checkpoint. Creating your own templates for your favorite models is the recommended approach.11. LoRAsDual LoRA Stack setup with two Lora Stack nodes feeding into a Lora Stack Apply. Supports both model-only and model+clip modes with up to 9 combined LoRA slots. Each slot has its own model path and strength controls.You can mix model-only LoRAs and LoRAs that also modify CLIP in the same stack — each slot independently selects its mode.12. Model PatcherA collection of 11 optional model modifications, each independently toggleable via a Fast Muter panel:ModelSamplingFlux — Flux-specific guidance parametersModelSamplingAuraFlow — AuraFlow sampling overrideDynamicThresholdingFull — CFG thresholding for better prompt adherencePerturbedAttentionGuidance (PAG) — Self-attention manipulation for more detailSelfAttentionGuidance (SAG) — Feature map attention enhancementDifferentialDiffusion — Mask-based selective denoisingCFGZeroStar — Alternative CFG guidance techniquePatchSageAttention — Memory-efficient attention (reduces VRAM usage)TorchCompileModel — JIT compilation for faster inferenceTeaCache — Token caching for speed improvementUNetTemporalAttentionMultiply — Temporal attention modificationMost of these are bypassed by default. Enable them one at a time to see their effect on your output.13. PuLID — FluxIdentity preservation using PuLID. Load a reference face photo and PuLID will guide the generation to maintain that person's facial features in the output. Uses pulid_flux_v0.9.0 with a configurable strength weight.14. PuLID — Flux NunchakuSame concept as PuLID Flux, but optimized for Nunchaku-quantized Flux models. Uses pulid_flux_v0.9.1 + EVA02_CLIP.15. Flux ReduxStyle transfer using Flux Redux. Load one or two style reference images and the workflow applies their visual style to your generation via CLIP Vision encoding and StyleModelApply. Supports blending two references at different strengths.16. PreprocessorImage preprocessing for ControlNet. Uses DepthAnything for depth map extraction. Only needed when the ControlNet group is active.This group is also needed when using Flux ControlNet LoRAs (like depth) or the DiffSynth Qwen LoRA — their sub-feature toggles are i2i (Flux Preproc) and i2i (DiffSynth: Qwen Lora) in the Initial Render group. You must activate the Preprocessor group manually when using either of these.17. ControlNetStructural conditioning with four toggleable modes:Standard ControlNet — xinsir union-promax (strength 0.75)Union Type — Select specific control type (depth, canny, etc.)Negative Zero — Zero-out negative conditioningDiffSynth Qwen/ZIT ControlNet — Alternative ControlNet using Z-Image-Turbo model (strength 0.65)Each mode has its own Mode Bridge toggle. You can use standard ControlNet for structure while also enabling negative zero-out, for example.Rendering (Groups 24–26)24. Initial RenderThe core generation step. Contains a component subgraph (42 internal nodes) that handles the actual sampling process.Sub-features controlled by individual toggles:Initial Render — The main txt2img or img2img sampling passSeed Enhancer — Adds noise variation to the seedNoise Injection — Additional noise patterns injected into the latent (strength 0.45)Detail Daemon — Micro-detail enhancement during samplingFlux Guidance — CFG control specifically for Flux modelsi2i (Denoise) — Standard img2img with configurable denoise strengthi2i (Flux Preproc) — Flux ControlNet LoRA pathway (e.g. depth) — requires the Preprocessor group to be activated manuallyi2i (DiffSynth: Qwen/ZIT) — Qwen-based img2img pathwayi2i (DiffSynth: Qwen Lora) — Qwen LoRA variant pathway — requires the Preprocessor group to be activated manuallyNegative Prompt — Enable/disable negative conditioningStop — Halt execution after this renderDefault sampler: euler / simple / 25 steps / cfg 3.5 / denoise 1.0 — configured via Smart Sampler Settings v2.25. Latent UpscaleSecond-pass latent-space upscaling. Takes the initial render's latent output, upscales it 1.25× with bicubic interpolation, and runs a second sampling pass using a ClownShark Sampler component (7 internal nodes from the RES4LYF pack). This is a more advanced sampler with detail boost, SDE, and sigma scaling options.Default sampler: dpmpp_2m / sgm_uniform / 36 steps / denoise 0.5. ClownShark sub-sampler: multistep/dpmpp_2m / beta / 11 steps / denoise 0.23.26. Initial Render — PreviewPreview and save checkpoint. Shows the generated image and optionally saves it with full metadata embedding (workflow JSON + generation data). Includes a Stop node so you can halt here before entering the post-processing pipeline in Row 2.This is the boundary between generation and post-processing. If you just want to generate and save without any post-processing, enable the Stop node here.Row 2 — Post-Processing PipelineThe second row handles everything after initial generation. A routing banner of ~39 ungrouped nodes sits above the row, resolving shared resources: latent dimensions, reference image, MODEL (with a 6-source priority chain), VAE, CLIP, conditioning, and string prompts.Each Row 2 group automatically picks up the image from whichever previous group was last active, so you can enable any combination and the pipeline chains them correctly.Refining1. Flux2/ZIT Refiner — 3rd PassA third-pass refinement using a dedicated checkpoint (darkBeast Klein2) via a component subgraph (17 internal nodes). Uses SamplerCustomAdvanced with wavelet color matching (strength 0.75) to preserve the original color palette while refining details. Low denoise (0.3) for subtle improvement without major changes.Face Swap2. Flux2: Face SwapDiffusion-based face replacement using a two-pass BFS architecture built entirely with Eclipse and core ComfyUI nodes — no third-party face swap package needed. Contains two component subgraphs — BFS_1ST (26 internal nodes) and BFS_2ND (22 internal nodes) — for progressive face re-generation.How it works:Smart Detection finds the face in the image using the Anzhc face segmentation modelThe face region is cropped and encoded to latentBFS_1ST re-generates the face region using a dedicated checkpoint (darkBeast Klein2) via SamplerCustomAdvanced with full denoise (1.0)BFS_2ND refines the result with a second sampling pass for seamless blendingAn Image Comparer shows before/after for quality checkingBecause it uses actual diffusion sampling rather than a face-swap model, the results respect the art style and lighting of the original image. The BFS subgraphs need a Flux2 model trained for face re-generation — darkBeast Klein2 works well, but any BFS-capable Flux2 checkpoint should work. There are also LoRAs that add BFS capability to a standard Flux2 model. If your main pipeline uses a different model, the BFS checkpoint must be loaded in the BFS model loader.Upscaling (Groups 3, 9, 10)3. Upscale ImageFirst upscale stage with three independently toggleable methods:Scale to Total Pixels — Resize to 2 megapixels using lanczos interpolationSmart Sharpen+ — 4-pass adaptive sharpening (strength 0.75)Upscale with Model — Neural network upscaler (4x AnimeSharp) — muted by defaultAn Image Comparer shows before/after.9. SeedVR2 UpscaleAI-powered upscaling using the SeedVR2 7B DiT diffusion model — a video upscaler repurposed for single images. Loads its own dedicated DiT model and VAE, processes in LAB color space for better color accuracy. Includes optional pre-resize and RAM cleanup controls. Bypassed by default (resource-heavy).10. Rescale ImageFinal size adjustment with three chained operations:Reinhard Color Match (strength 0.3) — Matches colors back to the original referenceBicubic Rescale at 1.25× with supersample enabled (on by default) — supersampling renders at a higher internal resolution then downscales for cleaner results. Can be turned off for a simpler resizeSmart Sharpen (2 passes) — Final sharpening passDetailing (Groups 4–8)All five detailer groups share an identical architecture built around a component subgraph (SEGS Detailer, 41 internal nodes each). Each detailer:Detects a specific region in the image (face, eye, mouth, etc.)Creates a precise mask using SAM2.1 + VITMatte for clean edgesInpaints just that region at a low denoise to enhance detail without changing the restCompares before/after so you can check the resultEach detailer can optionally load its own dedicated model (separate from the main pipeline), has its own LoRA stack, and its own sampler settings — making them fully independent. Sub-features (model loader, LoRAs, flux guidance, negative prompt, differential diffusion, CFG zero star) are individually toggleable via Mode Bridge controls.4. Detailer: FaceEnhances facial details. Uses Florence-2 VLM with "face" detection → SAM2.1 + VITMatte masking. Dedicated model: darkBeast Blitz6. Denoise: 0.2 (subtle refinement — just enough to sharpen features without changing the face).5. Detailer: EyeEnhances eye details. Same architecture, "eye" detection prompt. Denoise: 0.35 (slightly more aggressive than face to bring out iris detail and reflections).6. Detailer: MouthEnhances mouth/teeth details. "Mouth" detection prompt. Denoise: 0.4 (the most aggressive of the face-area detailers — teeth and lips benefit from more rework).7. Detailer: X-1Body region detailer using YOLO object detection instead of Florence-2. Denoise: 0.3.8. Detailer: X-2Specialized region detailer using YOLO detection. Denoise: 0.4.The detailers run in sequence: face → eye → mouth → X-1 → X-2. Each picks up the output of the previous one automatically. Disable any you don't need — the chain adapts.Watermarks & Save (Groups 11–13)11. Create Watermark — TextOverlays a text watermark ("© Eclipse") on the image. Configurable font, size, color, and position (default: bottom-right). Includes gradient effects (cyan→blue) and optional Drop Shadow + Outer Glow from LayerStyle.12. Create Watermark — LogoOverlays a logo image as a watermark. Loads a logo file, positions it bottom-right with blue gradient effects. Optional desaturation, resize, Drop Shadow, and Outer Glow.13. Save ImageThe final output node. Collects the finished image from the entire pipeline using a priority chain that checks all possible sources in reverse order:watermark_logo → watermark_text → rescale → seedvr2 → yolo2 → yolo1 → mouth → eye → face → upscale → bfs → refiner → init → ref_imageThis means it always saves the output from the last active processing stage, regardless of which groups are enabled. The image is saved with full embedded metadata — workflow JSON, generation data (all models, VAEs, and LoRAs collected from across the entire workflow, plus prompts, dimensions), and all relevant settings.This is the only group you should always keep active. Everything else is optional.Quick Start GuideSimplest Setup — Text to ImageMake sure the image input groups are bypassed (Image Load, Image Load from Folder, Input Video Frame)In the Prompt group, type your prompt in the Wildcard Processor text field (the main prompt input — set to fixed mode by default) and your negative prompt in the Negative Prompt fieldIn the Folder / Size group, set your desired image dimensionsMake sure Model Loader is active with your preferred checkpointMake sure Initial Render and Save Image are activeBypass everything else you don't needQueue the promptImage to ImageEnable Image Load and select your source imageEnable the Resize Image group so your image is resized to match the dimensions set in Folder / Size — this avoids issues with oversized images. You can skip this if your image already matches, but large images may cause problemsIn the Initial Render group, enable the i2i (Denoise) toggle and set your denoise strength (0.3–0.7 is typical)Queue the promptPost-Process an Existing Image (Skip Render)You can load any image and send it straight to Row 2 — bypassing the entire generation step:Enable Image Load and select your imageDisable the Initial Render switch to skip rendering entirelyEnable whichever Row 2 groups you want (Refiner, Detailer: Face, Upscale Image, Face Swap, etc.)Queue — the pipeline picks up your loaded image and runs it through the active post-processing chainThis is one of the most useful features of the workflow. You can bring in any image — from a different workflow, a different tool, or even a photograph — and run it through the full detailing, upscaling, and watermarking pipeline without generating anything.Adding Post-ProcessingGenerate your base image firstEnable the Row 2 groups you want (Refiner, Upscale Image, detailers, etc.)Re-queue — the pipeline will process through all active Row 2 groups automaticallyUsing DetailersEnable any detailer groups you want (Detailer: Face, Detailer: Eye, Detailer: Mouth, Detailer: X-1, Detailer: X-2)Each detailer auto-detects its target region — no manual masking neededCheck the Image Comparer in each group to verify the resultAdjust denoise strength if the changes are too subtle or too aggressiveTroubleshootingThe workflow stops halfway throughMany groups have a Stop toggle that halts execution after that group finishes. This is useful for checking intermediate results, but they are enabled by default in some groups. If the workflow stops unexpectedly, check the Stop toggles in these groups:Image source groups (Image Load, Image Load from Folder, Preview Cropped Image)Initial Render and Initial Render — PreviewEach detailer (Face, Eye, Mouth, X-1, X-2)Disable the Stop toggle in any group where you want execution to continue through to the end.If you queue the workflow and it seems to finish too early — before reaching Save Image — a Stop toggle is almost always the reason. Check the last group that produced output and disable its Stop switch.Toggles reset when activating a groupWhen you change a group's state (mute → active or bypass → active), all toggles in that group reset to their defaults — which means all enabled. This can turn on sub-features you didn't expect, including the Stop toggle. After activating a group, always review its toggle panel and disable anything you don't need.This is the most common source of confusion. If something behaves differently after you re-activate a group, check its toggles — they've all been reset to enabled.Custom Node Packages UsedPrimary (author's own):ComfyUI_Eclipse — The backbone of this workflow. Provides loaders, pipes, Set/Get routing, Mode Bridges, Mute/Bypass Repeaters, Smart Prompt, Smart Folder, Smart Detection, Smart LM Loader, Smart Sampler Settings, Save Images, Image Comparer, and many more.RES4LYF — ClownShark Sampler (advanced sampling with detail boost) — fork of ClownsharkBatwing/RES4LYFThird-party:Raffle — Random prompt generation from tag categoriespysssss Custom-Scripts — ShowText for prompt preview displayKJNodes — Image resize, PatchSageAttentionSeedVR2 VideoUpscaler — AI-powered upscalingNunchaku — Quantized model support and PuLID integrationImpact Pack — SEGSPreview for detailer visualizationLayerStyle — Drop shadow, outer glow, SAM2Ultra, MaskGrow, ImageAutoCrop, and moreLayerStyle Advance — Extended LayerStyle nodes (SAM2 Ultra V2, VITMatte)Advanced ControlNet — ACN_AdvancedControlNetApply_v2BiRefNet — Background removalVHS (VideoHelperSuite) — Video frame loadingIf you made it this far — you're a legend. Now go generate something beautiful. 🌒

Flux Metal Jacket

🚀 FMJ V13 est là !On continue d’enrichir la boîte à outils avec deux excellents modèles qui vont faire plaisir à la communauté.Nouveautés techniques :Ernie Image (Baidu)Un modèle Text-to-Image basé sur une architecture single-stream Diffusion Transformer (DiT) de 8B paramètres.Très bon en suivi de prompt (instruction following), rendu de texte dans l’image et cohérence structurée.Disponible en version standard (qualité maximale) et Turbo (seulement 8 étapes d’inférence pour une génération ultra-rapide).Idéal pour les prompts complexes, les mises en page précises et le rendu multilingue.Hunyuan3D (Tencent)L’un des meilleurs modèles Image-to-3D du moment (et Text-to-3D).Il produit des meshes de haute qualité avec une topologie propre, des géométries précises et des textures PBR exploitables directement en production.Parfait pour générer rapidement des assets 3D game-ready, des prototypes ou du contenu pour l’impression 3D.Ces deux ajouts renforcent considérablement les capacités de FMJ :→ Génération d’images de haute fidélité avec Ernie Image→ Passage fluide du 2D vers du 3D professionnel avec Hunyuan3DUn grand merci à l’équipe pour l’intégration rapide et soignée !Testez-les sans hésiter et partagez vos meilleurs résultats ici: https://www.facebook.com/groups/fluxmetaljacket, on a hâte de voir ce que vous allez créer 🔥#FMJ_V13 #ErnieImage #Hunyuan3D #AI #TextToImage #ImageTo3D

ComfyUI beginner friendly Ernie Image (Base or Turbo) GGUF Text-to-Image Workflow with Easy Prompt Saver by Sarcastic TOFU

This is a very simple workflow that helps you to save your Ernie Image (Base or Turbo) GGUF Text to Image Generation Data into a human readable .txt file. This will automatically get and write your metadata to the .txt file. You will find all the saved prompt files that it generated with the images inside the Archive (.Zip) that has the workflow. Also with the Image Saver Simple node used here you may embed the workflow itself with each saved image or save the image and workflow for your work separately.ERNIE-Image is an open-source text-to-image generation model developed by Baidu and officially released in mid-April 2026. Built on a compact 8-billion parameter Diffusion Transformer architecture, it is designed to run efficiently on consumer-grade hardware while delivering state-of-the-art performance in multilingual text rendering and complex layout generation. Alongside the standard model, Baidu introduced ERNIE-Image-Turbo, a distilled variant optimized using Distribution Matching Distillation and reinforcement learning. While the base model typically requires 50 inference steps to reach peak quality, the Turbo version can generate high-fidelity images in just 8 steps, providing a massive speed advantage with minimal loss in detail. Both models are released under the Apache 2.0 license, making them highly accessible for developers and creators looking for a balance between speed, control, and precise instruction following.You can download all of your necessary model files used in this workflow from HuggingFace and LORA from CivitAI (Details are mentioned below). Make sure you have latest enough ComfyUI installation and install any necessary nodes for for this workflow using ComfyUI manager and place the correct files in correct places. Also check out my other workflows for SD 1.5 + SDXL 1.0, Pony, WAN 2.1, WAN 2.2, MagicWAN Image v2, QWEN, HunyuanImage-2.1, HiDream, KREA, Chroma, AuraFlow, NoobAI, Illustrious, Lumina2, Z-Image Turbo, Flux.2 Klein 9B & 4B, Flux.1 Dev and Kandinsky Image 5 Lite (T2I & I2I) models. Feel free to toss some yellow Buzz on stuffs you like.How to use this -#1. Just select your Ernie Image (Base or Turbo) GGUF model files first and#2. set your desired image dimensions to start#3. then input your desired image prompt.#4. select how many images you want (Change the number besides the "Run" button)#5. select image sampling methods, CFG, steps etc. settings#6. finally press the run button to generate. That's it..** LORA usage for this workflow is optional you can use it without any LORAs, use with 1 or 2 or any other number of LORAs, to add new LORAs press the L button on top to lunch LORA Manager on a new tab find your LORA and if you want to use that LORA just click the upward Kite button.Required Files===============### Download Link for Ernie Image GGUF Models -++++++++++++++++++++++++++++++++++++++++++++++++https://huggingface.co/unsloth/ERNIE-Image-GGUF/resolve/main/ernie-image-Q4_K_M.ggufhttps://huggingface.co/unsloth/ERNIE-Image-Turbo-GGUF/resolve/main/ernie-image-turbo-Q4_K_M.gguf### Download Link for Ernie Image FP8 Models (Optional) -++++++++++++++++++++++++++++++++++++++++++++++++++++++++++https://huggingface.co/Bedovyy/ERNIE-Image-Quantized/resolve/main/ernie-image-fp8.safetensorshttps://huggingface.co/Bedovyy/ERNIE-Image-Quantized/resolve/main/ernie-image-turbo-fp8.safetensorsYou can use these if you have trouble running both the full regular models & GGUF models but your hardware can handle these FP8 models (12GB or higher VRAM)### Download Link for Ernie Image Text Encoder -+++++++++++++++++++++++++++++++++++++++++++++++++https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/text_encoders/ministral-3-3b.safetensors### Download Link for Ernie Image VAE -++++++++++++++++++++++++++++++++++++++++https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/vae/flux2-vae.safetensorsYes! as you can see it is just your Flux.2 VAE, if you have already downloded this for Flux.2 Klein (4B or 9B) you can skip downloading it again and just use the one you already used.### Download Link for Ernie Image Prompt Enhancer (Optional) -+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++https://huggingface.co/Bedovyy/ERNIE-Image-Quantized/resolve/main/text_encoders/ernie-image-prompt-enhancer-fp8.safetensorsI have skipped using the Prompt Enhancer in this workflow for simpler and faster generation but if you are someone who likes to experiment with existing workflows like me and have better hardware that can handle FP8 models (12GB or higher VRAM) you can tweak and incorporate this with Ernie Image (Base or Turbo) FP8 modelLORA used -++++++++++++ERNIE-Image Turbo Lora -https://civitai.red/models/2551180/ernie-image-turbo-lora?modelVersionId=2867032This Lora extracted from Ernie Image Turbo can reduce the number of steps to 12 for faster generation on basic (8GB or 12GB VRAM) GPUs when used with Ernie Image Base model.. the resulting image sometimes can produce slighlty degraded outputs. You can skip this if you are using the Ernie Image Turbo model or if you want to use the standard 30 to 50 steps, CFG 5 to 7.5 with Ernie Image Base model for your desired output.

Rebels Ernie Image Base + Z-Image Base

🌟 Ernie Image BASE + Z-Image BASE Refiner 🌟Get the incredible detail of Ernie without the visual artifacts!🛑 The Problem:While Ernie Image BASE is incredibly powerful and produces stunning compositions, it has a known tendency to degrade otherwise great generations with harsh blowouts (overexposed highlights) and heavy stress lines.✨ The Solution:This workflow uses Z-Image BASE as a dedicated refiner to clean up the output. By running the Ernie generation through a Z-Image pass, it completely smooths out those harsh artifacts, leaving you with a polished, pristine final image.⚙️ How It Works (The Sweet Spot):This setup is specifically tuned to fix artifacts while strictly preserving your original composition. It is also highly optimized, running smoothly even on 8GB VRAM setups like an RTX 3070:• Base Pass: Runs [Insert Number] steps on Ernie Image BASE to establish the core image.• Refiner Pass: Runs [Insert Number] steps on Z-Image BASE. • Denoising Strength: Set to exactly 0.35.💡 Why 0.35 Strength?This low denoising strength is the magic number. It ensures the original Ernie composition is preserved as much as possible, while giving Z-Image just enough room to act as a cleanup crew for the blown-out lighting and stress lines.📦 Required Models:To get this working, you will need to have the following models installed:Ernie Image BASEZ-Image BASE

Rebels EIT + ZIT refiner

🌟 Ernie Image Turbo + Z-Image-Turbo (Refiner) 🌟Get the blazing speed of Ernie without the visual artifacts!🛑 The Problem:While Ernie Image Turbo is incredibly powerful and fast, it has a known, frustrating tendency to degrade otherwise great generations with harsh blowouts (overexposed highlights) and heavy stress lines.✨ The Solution:This workflow uses Z-Image-Turbo as a dedicated refiner to clean up the output. By running the Ernie generation through a quick Z-Image pass, it completely smooths out those harsh artifacts, leaving you with a polished, pristine final image.⚙️ How It Works (The Sweet Spot):This setup is specifically tuned to fix artifacts while strictly preserving your original composition:• Base Pass: Runs 8 steps on Ernie Image Turbo to establish the core image.• Refiner Pass: Runs 4 steps on Z-Image Turbo. • Denoising Strength: Set to exactly 0.35.💡 Why 0.35 Strength?This low denoising strength is the magic number. It ensures the original Ernie composition is preserved as much as possible, while giving Z-Image just enough room to act as a cleanup crew for the blown-out lighting and stress lines.📦 Required Models:To get this working, you will need to have the following models installed:Ernie Image TurboZ-Image Turbo

Ernie Image + Prompt Enhancer Workflow

Ernie-Image-HQ

Ernie-Image Simple Advanced Workflow for HQ OutputsGGUF: https://huggingface.co/unsloth/ERNIE-Image-Turbo-GGUF/tree/mainCLIP (TE): https://huggingface.co/Comfy-Org/ERNIE-Image/blob/main/text_encoders/ministral-3-3b.safetensorsGGUF CLIP (TE): https://huggingface.co/Abiray/Huihui-Ministral-3B-Instruct-2512-abliterated-GGUF/tree/mainVAE: https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/blob/main/full_encoder_small_decoder.safetensors

ERNIE Image NVFP4 (With Optional Turbo LoRA, Prompt Enhance, 2nd-Pass)

Hello there!So, as the title says, this is an ERNIE Image NVFP4 with Optional Turbo LoRA, Prompt Enhance, 2nd-Pass Workflow. You can also use other ERNIE models (base, turbo) and any other ERNIE LoRA. If you don't want to use the prompt enhancer, you can disable it too. The workflow includes instructions.Tested with/in:Intel Core Ultra 7, RTX 5060 Ti 16GB, 32GB DDR5ComfyUI Portable (Windows) 0.19.2Resources:Text Encodersernie-image-prompt-enhancer.safetensors➡️ https://huggingface.co/Comfy-Org/ERNIE-Image/blob/main/text_encoders/ernie-image-prompt-enhancer.safetensorsministral-3-3b.safetensors➡️ https://huggingface.co/Comfy-Org/ERNIE-Image/blob/main/text_encoders/ministral-3-3b.safetensorsDiffusion Modelsernie-image-nvfp4.safetensors➡️ https://huggingface.co/Starnodes/quants/blob/main/ernie-image-nvfp4.safetensorsVAEflux2-vae.safetensors ➡️ https://huggingface.co/Comfy-Org/ERNIE-Image/blob/main/vae/flux2-vae.safetensorsLoRA (OPTIONAL)enrnie-turbo-lora ➡️ https://civitai.com/models/2551180/ernie-image-turbo-loraCustom NodesRES4LYF ➡️ https://github.com/ClownsharkBatwing/RES4LYFFiles folders: 📂 ComfyUI/ ├── 📂 models/ │ ├── 📂 diffusion_models/ │ │ ├── ernie-image-nvfp4.safetensors │ ├── 📂 loras/ │ │ ├── extracted_lora.safetensors │ ├── 📂 text_encoders/ │ │ ├── ernie-image-prompt-enhancer.safetensors │ │ └── ministral-3-3b.safetensors │ └── 📂 vae/ │ └── flux2-vae.safetensorsAny Buzz donation is welcome! ❤️ Thanks!

Ernie GGUF Switcher: Turbo-Charged Flux Efficiency

This workflow is designed for creators who demand high-fidelity results without the heavy hardware tax. By leveraging the Ernie-Image-Turbo model in GGUF format, this setup provides a streamlined, fast, and VRAM-efficient path to professional-grade AI imagery.Key FeaturesGGUF Optimization: Optimized for lower VRAM usage via the UnetLoaderGGUF node, allowing high-end generation on mid-range hardware.Turbo Performance: Uses the res_2s sampler with only 8 steps, delivering hyper-realistic results in a fraction of the time.High-Resolution Ready: Defaulted to a 720x1024 portrait ratio, perfect for cinematic character shots and detailed textures.Advanced CLIP Integration: Utilizes the ministral-3-3b CLIP model for superior prompt adherence and nuanced understanding of complex instructions.Components & NodesModel: ernie-image-turbo-Q8_0.ggufVAE: flux2-vae.safetensorsCLIP: ministral-3-3b.safetensorsSampler: KSampler configured for speed (res_2s / simple scheduler).How to UseEnsure you have the ComfyUI-GGUF custom nodes installed.Place your GGUF model in the models/unet folder and CLIP/VAE in their respective directories.Input your prompt into the PrimitiveStringMultiline node and hit generate to see the "Turbo" efficiency in action.

Ernie Resolution

Ernie Resolution (ComfyUI)Utility node for Baidu ERNIE-Image with preset recommended resolutions. https://huggingface.co/baidu/ERNIE-ImageSelect a resolution and get ready-to-use latent, along with width and height outputs.Includes official aspect ratios for optimal quality and stability.Installation:Place the Ernie Resolution.py file into your ComfyUI/custom_nodes/ folder and restart ComfyUI.

Simple Ernie Turbo GGUF Workflow.

*If you found this useful, please give it a like!A very simple workflow for using Ernie GGUF models checkpoints in ComfyUI.Ready for use Ernie Turbo GGUFDownload Ernie Turbo GGUF checkpoints: https://huggingface.co/unsloth/ERNIE-Image-Turbo-GGUF/tree/mainDownload Ernie GGUF checkpoints: https://huggingface.co/unsloth/ERNIE-Image-GGUF/tree/main

ERNIE-AIO-Turbo t2i v.1 powered by Ollama

Easy to use Ernie-AIO text to image workflow, with SeedVR upscaler.

Lonecat's Ernie Workflow (Base & Turbo)

Note: If you have ANY issues with nodes not downloading, read the notes or reach out. There's nothing that special about any of them that aren't core modules. Most issues can be resolved by updating everything⛔⚠️🛑✋ Read the notes completly before using. Most common install and node problems are listed in the directionsInstagram: https://www.instagram.com/synth.studio.models/Buy me a☕ https://ko-fi.com/lonecatoneThis represents many hours of work. If you enjoy it, please 👍like, 💬 comment , and feel free to ⚡tip 😉也有中文说明A complete user and troubleshooting guide inclusing a list of common error messages can be found here https://civitai.com/articles/24678/how-to-use-my-workflows-a-comprehensive-breakdownPrompt assistDetailersPower LoRa loader (for whenever people start making LoRas)Qwen Prompt enhancementUltimate Upscaler for prescaling and hi res fixMassive Post Production Suite📷 Optical realismJPEG ScrubberSmart Noise scrubberInline crop tool for easy image adjustmentsFolder & Metadata save optionsFull tutorial available.也有中文说明

ERNIE Image basic workflow (Base + Turbo)

Workflow descriptionThis is a very basic ERNIE workflow for ComfyUI, available in two variants: base and turbo.No prompt enhancers, no detailers, no extra nodes. Just the core setup to generate images.The goal is simple: give you a clean, stable foundation that you can expand yourself. Add your own prompt enhancers, detailers, LoRAs, or refinements on top of this.If you are getting started with ERNIE or want a minimal workflow without noise, this is exactly that.Access All My WorkflowsWant access to all my workflows, LoRA dataset training setups, and more?Check: https://huslyo123.carrd.co/You will find all details there. Fanvue members get automatic access to my private Discord.Models and setupDiffusion modelsBase:https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/diffusion_models/ernie-image.safetensorsTurbo:https://huggingface.co/Comfy-Org/ERNIE-Image/blob/main/diffusion_models/ernie-image-turbo.safetensorsText encoderhttps://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/text_encoders/ministral-3-3b.safetensorsVAEhttps://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/vae/flux2-vae.safetensorsFolder structureComfyUI/ └── models/ ├── diffusion_models/ │ ├── ernie-image.safetensors │ └── ernie-image-turbo.safetensors ├── text_encoders/ │ └── ministral-3-3b.safetensors └── vae/ └── flux2-vae.safetensorsRecommended settingsBase modelSteps: 50CFG: 4.0Sampler: EulerTurbo modelSteps: 8CFG: 1.0Sampler: res_multistepSupported resolutions1024x1024848x12641264x848768x1376896x12001376x7681200x896More infohttps://huggingface.co/baidu/ERNIE-Imagehttps://huggingface.co/baidu/ERNIE-Image-TurboClip loader error? Update ComfyUI to the latest release. Reboot.Update both .bat and manager if problem still exists. Reboot.Access All My WorkflowsWant access to all my workflows, LoRA dataset training setups, and more?Check: https://huslyo123.carrd.co/You will find all details there. Fanvue members get automatic access to my private Discord.