This chapter selects existing workflows with instructions and public examples for character consistency, product images, local editing and upscaling. They are starting points for reuse; results still depend on inputs and runtime conditions. Resources and original workflow configurations were checked on October 31, 2026; local image generation has not been acceptance-tested.
\nWorkflow resource entry points
\n| Source | What it provides | How to use it |
|---|---|---|
| Official ComfyUI templates | Loadable workflows, model links and instructions | Start with a template, obtain its models and run it |
| Official Comfy workflow directory | Workflows organized by product images, lighting, multiple character views, local editing and upscaling | Find candidates by task |
| ComfyUI maintainer's examples | SDXL, image-to-image, inpainting, LoRA, ControlNet and upscaling examples | Download and load original images containing workflow information |
| RunComfy workflow library | Case descriptions, workflow JSON and preconfigured environments | Follow along in the cloud or reproduce locally using the dependency instructions |
Four task types at a glance
\n| Task | Selected existing workflow | Main inputs and conditions |
|---|---|---|
| Character consistency | Author's IP-Adapter FaceID Plus v2 example | Clear face reference; matching SD base model, FaceID, LoRA, vision encoder and InsightFace; research use |
| Product images | RunComfy Product Relighting, plus a separate basic local IC-Light example | Clear product image, optional background and lighting mask; the full workflow and basic example have separate dependencies |
| Local editing | Official ComfyUI Inpainting example | Original image, edit mask and specified inpainting model; native nodes |
| Upscaling | Official ComfyUI model-upscaling workflow with ESRGAN or Real-ESRGAN | Original image and upscaling model; the basic workflow needs no SD checkpoint, prompt or seed |
Conditions and limits
\nUse one face reference to practice retaining the same person's facial features across scenes. Begin with a clear, unobstructed, centered image of one face. This workflow mainly constrains facial identity; it does not also guarantee consistent clothing, body shape, pose and setting. The model author notes that identity consistency is imperfect.
\nThe FaceID model card restricts use to research. Do not treat this example as a ready-to-use commercial portrait solution. The plugin author has placed the project in maintenance mode; record working ComfyUI and plugin versions and retain their files. Uses and capabilities, plugin installation and maintenance
\nResources and workflow files
\nReuse steps
\nip-adapter-faceid-plusv2_sd15.bin, its paired ip-adapter-faceid-plusv2_sd15_lora.safetensors, and the README's ViT-H vision encoder. Put them in models/ipadapter, models/loras and models/clip_vision, respectively; put the checkpoint in models/checkpoints. Prepare InsightFace's face models as the plugin instructs.Conditions and limits
\nSuitable for changing the presentation environment or lighting of an existing product photo. The full example is Andrea Baioni's Product Relighting; its RunComfy page includes a JSON download, illustrations and the author's video. A clear subject image is required; background and light masks are optional. These conditions and the switches below come from the case instructions.
\nThese workflows regenerate or harmonize an image. Compare the output with the original product's shape, color, packaging text, logos and small components. Restore exact text or logos from original assets during subsequent refinement when needed.
\nResources and workflow files
\nReuse the complete product-image workflow
\nRequirements for the basic local example
\nKijai's original PNG stores the SD 1.5-branch photon_v1.safetensors, iclight_sd15_fc.safetensors and nodes including ImageResizeKJ. Prepare that checkpoint, ComfyUI-IC-Light and the required KJNodes. Put the IC-Light model in the appropriate directory under models/unet according to the node documentation and select it in the loader. Replace the sample input, describe the subject and lighting, and start with the original approximately 512-pixel configuration. Compare results separately if using another checkpoint.
The IC-Light author also notes that BRIA RMBG 1.4 in the original demo is restricted to noncommercial use. If a workflow uses a background-removal model, check that component's conditions for product use; the main IC-Light code license alone does not determine the whole workflow's terms.
" }, { "id": "course-image-workflow-reuse-section-3", "title": "Local editing: official ComfyUI Inpainting", "url": "https://ai-knowledge-map.com/en/software/image-creation/image-workflow-reuse/#course-image-workflow-reuse-section-3", "html": "Conditions and limits
\nSuitable when the overall image is satisfactory and only a region such as a hat, clothing texture or background clutter needs changing. You need the original image and an edit mask. For basic practice, use the officially specified 512-inpainting-ema.safetensors with an approximately 512-pixel input; do not add third-party nodes initially.
Although the original workflow filename contains sd1.5, its loader specifies the inpainting file above. Follow the node and model requirements rather than inferring the model version from an image filename.
\nResources and workflow files
\nReuse steps
\nComfyUI/models/checkpoints and select it in Load Checkpoint. Keep the template's sampling settings initially.Conditions and limits
\nSuitable for photos or illustrations whose composition is settled but need larger pixel dimensions. ComfyUI's native model-upscaling workflow needs no additional SD checkpoint, prompt, LoRA or seed. Inspect text, outlines, faces and textures afterward; more pixels do not guarantee recovery of real details absent from the original.
\nResources and workflow files
\nModel selection reference
\n| Image and goal | Models to compare |
|---|---|
| General image upscaling | RealESRGAN_x4plus or RealESRGAN_x2plus |
| Anime and illustrations | RealESRGAN_x4plus_anime_6B |
| Lower runtime and VRAM use | realesr-general-x4v3; its author notes more limited deblurring and denoising |
Reuse steps
\nComfyUI/models/upscale_models and select it in UpscaleModelLoader. The saved RealESR/4x-ESRGAN.pth path is the author's directory record; reselect your actual local file.Run each workflow with its original configuration before replacing inputs. Models, LoRAs, nodes and assets are usually not all bundled inside one JSON file. Without an explicit source, do not assume a universal minimum VRAM requirement: local needs depend on models, dimensions, precision, simultaneously loaded components and software versions. Use RunComfy or other cloud environments according to their current configurations and fees.
\nFor official templates, use Workflow → Browse Workflow Templates and obtain models indicated as missing. Load author-shared workflows from JSON or original PNG, then check custom nodes. See template instructions and missing-node and model troubleshooting.
\nKeep these records for further changes or comparisons:
\n| Record | Content |
|---|---|
| Workflow source | Page, filename, download date and identifiable version or commit |
| Runtime configuration | ComfyUI and custom-node versions, models and companion files, key parameters |
| Inputs and goal | References, masks, prompts, content to retain and changes to make |
| Checks | Whether identity, product details, edit boundaries or upscaled textures meet requirements |
| Actual effort | Runtime, execution errors and manual correction |
Matching seeds and parameters still do not guarantee identical outputs across environments; see the Diffusers guide for reproducibility conditions. Compare candidates using the same assets to determine which workflow fits the current task.
" } ] }