comfyui-FVMtools Introduction
comfyui-FVMtools is a collection of custom nodes for ComfyUI that helps you build images with recognizable characters, coordinated colors, and detailed outfits. It also includes tools for placing prompts in specific parts of an image and repairing text that a generated image renders poorly.
You can use its tools separately or combine them in one workflow. For example, you might choose a palette, generate an outfit description using that palette, create a group portrait, and then refine each person with a different reference image, prompt, and LoRA.
The main tool groups are:
- Face Tools: Find and match people in an image, create masks for specific areas, and refine each person separately.
- Color Tools: Generate a palette or extract one from an image, then turn color placeholders in prompts into named colors.
- Fashion Tools: Build varied outfit descriptions from themed clothing lists.
- JB (JSON Builder): Assemble structured scene, location, and outfit prompts, or turn them into natural-language descriptions.
- K2 Lab: Assign prompts and LoRAs to drawn regions of a Krea 2 image.
- Sign Tools: Identify lettering that looks implausible and help re-render it.
How comfyui-FVMtools Works
ComfyUI workflows connect nodes like steps in a production process. In comfyui-FVMtools, one node can identify where an edit should happen, while another decides what to generate there.
For character refinement, Person Selector Multi examines an image and compares its detected faces with your reference images. It produces PERSON_DATA: a package of match information and masks that tells Person Detailer which image areas belong to which people. A mask is like a stencil: it marks the region that can be changed while helping preserve the rest of the picture. Person Detailer then works through the enabled people one at a time, applying each person’s chosen prompt and LoRA to the appropriate area.
For color and clothing, the process is simpler. An outfit description can contain placeholders such as #color1# and #accent#. Prompt Color Replace fills those placeholders using colors from a generated palette or an image you supplied. For instance:
Before: #color1# silk blouse, #accent# leather shoes
After: ivory silk blouse, burgundy leather shoes
The exact colors depend on the palette you use. This approach lets you change the color scheme without rewriting every outfit prompt.
Some tools provide additional guidance:
- A depth map represents which parts of a scene appear nearer or farther away. Face Tools can use it to improve masks and decide which person to refine first when people overlap.
- A ControlNet can help Person Detailer retain depth or pose while it refines a person.
- K2 Lab lets you draw boxes over an image and give each box its own prompt and LoRA. This is useful when different characters or objects need different instructions in a single Krea 2 generation pass.
comfyui-FVMtools Features
Face matching and per-person refinement
Person Selector (Match) compares faces in an image with one or more reference images and reports the best similarity score. You can choose whether it also produces a face, head, or body mask. Its threshold controls how similar a face must be to count as a match: raising it makes matching more selective; lowering it allows less similar faces through.
Person Selector Multi handles scenes with several people. It supports up to 10 reference slots and assigns each detected face to at most one reference. Its annotated preview helps you check those assignments before refining the image.
When faces look alike, you can adjust match_weights to consider more than facial identity. The available signals are face, hair, head appearance, and outfit colors. For example, 100/0/0/0 uses facial identity alone, while 60/20/20/0 also considers hair and head appearance. If you connect outfit palette previews, the node can use clothing colors and display palette swatches above matched faces. See the Matching Guide for tuning advice.
The selector can produce masks for the face, head, body, hair, facial skin, eyes, mouth, neck, and accessories. A connected segm_detector can supply separate body regions for each person, which is particularly useful in crowded scenes. An optional depth map can help distinguish overlapping regions and establish a back-to-front processing order.
Person Detailer uses the selector’s PERSON_DATA to refine up to five reference-matched people, plus an optional generic slot for other faces. Each slot can have its own LoRA, prompt, LoRA strength, and mask choice. You could, for example, give one character a face-only correction and another a head-level edit that includes the hair. Its denoise setting controls how strongly the inpainted area can change; use gentler values when you want to retain more of the original image. Mask blending and expansion settings control how the edited area meets its surroundings.
Inpaint Options gives you finer control over each slot’s mask and number of passes. For a two-round refinement, a progression such as 0.5|0.3 for denoise and 6|4 for steps applies a stronger first pass followed by a gentler one. Detail Daemon Options adjusts when detail-preserving behavior is applied during sampling.
Person Detailer ControlNet adds depth or pose guidance to the same per-person process. It reduces the effective ControlNet strength as denoise increases, helping limit guidance-related artifacts. Pose maps are smoothed to reduce hard lines appearing in the result.
Person Data Refiner is for workflows that upscale an image between detailing passes. It detects faces again at the new resolution, keeps their original reference assignments by position, and regenerates masks that fit the larger image. This avoids relying on masks made for the smaller version.
The detailer caches LoRA files and reuses repeated LoRA-and-prompt combinations during batch processing, reducing redundant loading and text encoding.
Color palettes and outfit prompts
Color Palette Generator creates palettes of 2–8 named colors from a selection of 161 color names. You can choose a harmony such as analogous, complementary, or monochromatic, and select from 14 style presets, including beach, urban streetwear, evening gala, and winter cozy. Sliders for vibrancy, contrast, warmth, and neutral tones let you steer the mood. A seed lets you reproduce a palette you liked.
Palette From Image samples a reference image instead. Its dominant mode emphasizes common colors, vibrant favors stronger colors, and fashion_aware is intended to prioritize clothing colors over skin and background. You can also sample only part of an image, such as its upper or lower half, and use the preview to inspect the result.
Prompt Color Replace turns tags into palette colors. It supports #color1# through #color8# (or #c1# through #c8#) and role-based tags such as #primary#, #secondary#, #accent#, #neutral#, and #metallic#. Tags are case-insensitive. If a prompt asks for a color that the palette does not contain, the node uses its configurable fallback_color.
Outfit Generator selects clothing and accessories from 40 built-in themed outfit sets. You can adjust formality and coverage, switch garment categories on or off, and control the chance of prints or text decorations. Its seed makes a selection reproducible. The Edit List button lets you change outfit-list entries in the browser; those changes take effect on the next queued run. The generated prompt can then pass through Prompt Color Replace for coordinated colors. The Outfit Guide explains themes, fabrics, and list customization.
Structured prompts, regions, and lettering
JB (JSON Builder) provides five nodes for assembling prompts:
- JB · Outfit Block and JB · Location Block make ready-to-use prompt sections with palette and list-editing controls.
- JB · Builder lets you arrange
key:valueentries into a nested prompt. - JB · Stitcher combines multiple prompt fragments.
- JB · Extractor retrieves one section from a larger prompt.
JB can output structured text for a text-encoding node. For Krea 2 / Qwen3-VL text encoders, its natural output presents the information as plain prose, while sentences retains the meaning of each prompt key in readable sentences. This can help when a text encoder treats JSON syntax as text rather than following its structure. The JB reference workflow and JB guide for Krea 2 show the available approaches.
StructPromptMaker (SMP) is an earlier, still-functional 13-node prompt pipeline. For most new structured-prompt workflows, the original README recommends JB; the SMP reference workflow remains available for existing setups.
K2 Lab contains 22 nodes for Krea 2 regional control. Its visual region builder lets you draw boxes over a recent render in a separate editor window. You can assign each box a prompt and one or more LoRAs, save or load a K2Lab project, and use tools for regional editing and face refinement. Region layouts rescale when the aspect ratio changes. See the K2 Lab guide.
Sign Tools help with generated lettering on signs, labels, clothing, posters, screens, books, license plates, paper, and graffiti. The tools identify likely text areas, propose scene-appropriate wording using a vision model in LM Studio, and re-render the chosen text with guidance based on letter shapes placed over the sign. Areas too small to read can instead be softened into plausible out-of-focus text. An optional OCR backend can improve detection of lettering that looks like text but cannot be read. The Sign Guide covers the workflow and its limitations.
comfyui-FVMtools Models
These are supporting models and model types, not separate versions of comfyui-FVMtools. Which ones you need depends on the nodes in your workflow.
| Model or model type | What it does | When it matters |
|---|---|---|
InsightFace buffalo_l | Detects faces and creates face-identity representations for comparison. | Face matching in Person Selector nodes. |
BiSeNet parsing_bisenet.pth | Separates facial and head areas into regions such as skin and hair. | Creating detailed face, head, and auxiliary masks. |
| SAM2.1 | Provides segmentation that can help define a body region. | Optional body-mask workflows using a SAM model. |
| Instance segmentation detector | Produces separate body regions for individual people. | Recommended by the original README for clearer, non-overlapping masks in multi-person scenes. |
| Depth model, such as Depth Anything V2 | Produces a depth map showing relative distance in the scene. | Optional mask refinement, overlap handling, and rendering order. |
| Depth or pose ControlNet | Guides an inpainted result toward the original depth or pose. | Person Detailer ControlNet workflows. |
| Your base image model and LoRAs | Generate and personalize the refined image areas. | Person Detailer and supported K2 Lab workflows. |
The original README recommends sam2.1_hiera_large.pt when using SAM2.1. It also notes that Person Detailer can detect Z-Image Turbo (Lumina2) and automatically convert certain LoRA attention weights to a compatible format when needed; you do not need to perform that conversion manually.
For multi-person body masks, the original README favors a connected instance segmentation detector over SAM-only segmentation. See the Person Detailer specification for a more technical account of the face workflow.
Troubleshooting comfyui-FVMtools
A person is matched to the wrong reference
- Check the annotated output from Person Selector Multi to see which face received each reference.
- Try a clearer reference image that shows the intended person’s face.
- If the faces are similar, adjust
match_weightsto consider hair and head appearance. If clothing colors distinguish the people, connectoutfit_palettesand include outfit matching. - Review
auto_thresholdor the manualthreshold; a stricter threshold can reject weak matches. - Consult the Matching Guide for more detailed examples.
A body edit affects another person or an overlapping object
- Inspect the selector’s body masks and preview before running the detailer.
- For several people, connect a
segm_detectorand usebody_mask_mode: autoso the node can use separate person segments. - If people or objects overlap, connect a depth map and check the rendering order in the preview.
- If boundaries remain rough, adjust mask blur or the available depth-edge settings, then compare the preview again before inpainting.
Masks no longer line up after upscaling
- Keep the original
PERSON_DATAfrom Person Selector Multi. - Send the upscaled images and that data to Person Data Refiner.
- Use the refiner’s new
PERSON_DATAfor the high-resolution Person Detailer pass. - Make sure the upscaled image batch has the same number of images as the original batch.
The refined area changes too much or has a visible border
- Lower Person Detailer’s
denoisefor a gentler edit. - Check whether the slot is using the intended mask type. A face mask changes less area than a head or body mask.
- Adjust
mask_blend_pixelsto soften the transition, or reviewmask_expand_pixelsif the edited area misses part of the target. - If using ControlNet, compare the result with depth or pose guidance to see whether it better preserves the original structure.
A color tag remains unresolved or becomes the fallback color
- Connect a palette’s
palette_stringoutput to Prompt Color Replace. - Confirm the tag is supported, such as
#color1#or#accent#. - Check that the palette contains enough colors:
#color7#cannot select a seventh color from a five-color palette. - Inspect
replacements_logand setfallback_colorto a suitable alternative where needed.
Outfit text or sign lettering looks implausible
For clothing generated by Outfit Generator, try its text_mode: descriptive to request a general decoration instead of exact words, or off to omit text decorations. For lettering already present in an image, use the Sign Tools workflow and the Sign Guide. Very small text may be better treated as out-of-focus lettering than forced into readable words.
Frequently asked questions
Do you need to use every tool group?
No. You can use color, fashion, face, JB, K2 Lab, and Sign Tools in separate workflows.
Can you give two people different LoRAs?
Yes. Person Detailer has five reference slots with separate LoRA and prompt controls, plus an optional generic slot.
Can you reuse a palette from an existing image?
Yes. Use Palette From Image, then feed its palette string into Prompt Color Replace.
What should you use after enlarging an image?
Use Person Data Refiner to regenerate masks at the larger resolution before another detailing pass.
Is SMP obsolete?
No. It remains functional, although the original README describes JB as the preferred suite for most new structured-prompt uses.
Learn More about comfyui-FVMtools
- Matching Guide: Face, hair, head, and outfit matching; useful when several characters look similar.
- Color Guide: Palette controls, style presets, and examples of colorized prompts.
- Outfit Guide: Outfit themes, fabrics, layering, custom lists, and text on clothing.
- JB reference workflow and JB for Krea 2: Examples of structured and natural-language prompt assembly.
- SMP reference workflow: An example for the earlier structured-prompt pipeline.
- K2 Lab guide: Visual regions, regional LoRAs, editing, and projects for Krea 2.
- Sign Guide: Finding and repairing generated lettering.
- Person Detailer specification: Deeper technical detail about matching, masks, and per-person inpainting.
