# ComfyUI Visual Quality Bakeoff V1

Date: 2026-07-14

## Goal

Find a ComfyUI checkpoint and LoRA recipe that can reproduce the benchmark's
production readability without copying its branded characters or exact drawing
identity. The required grammar is simple full-body silhouettes, bold uniform lines,
flat limited colors, readable exaggerated expressions, and registered whole-character
state changes.

## Downloaded Candidates

- Illustrious XL v2.0 checkpoint
- Juukyuu Flat Colors Illustrious LoRA
- Emotional Flat Illustration SDXL LoRA
- NTC Cartoon Slider SDXL LoRA
- BACKTOON SDXL background LoRA

Every installed file was hash-verified and recorded in
`config/comfyui_visual_model_candidates_v1.json`. Candidates with restrictive or
unclear commercial-image permissions were excluded before testing.

## Neutral Character Result

No public recipe passed.

- Animagine plus iroiro remains the least unstable technical baseline, but it cannot
  lock a prompt-only character and tends toward generic proportions and faces.
- Illustrious plus Juukyuu produced duplicate and incomplete figures despite `solo`
  and `multiple characters` constraints.
- Emotional Flat produced a clean low-texture surface but lost the requested body,
  skin, and costume design.
- Cartoon Slider made the silhouette rounder and lines bolder, but eyes and identity
  remained unstable. It is a style-strength experiment, not an identity solution.

Complex identity elements such as the gat and hanbok failed more severely than a
plain office character. Prompt-only input therefore requires an approved master
before any state generation.

## Registered Three-State Result

The canonical Godfortune reference was tested without face cropping or face inpaint.
IPAdapter Plus supplied identity reference and Xinsir OpenPose supplied body structure.

At IPAdapter `0.78` and OpenPose `0.90`, resemblance was partial but the extreme
arm-up pose was ignored. At IPAdapter `0.65` and OpenPose `1.35`, both action poses
became readable, but the face, hair, hat, body proportions, and costume changed into
a different character.

This is the current central failure: reference strength and pose strength trade off
against each other instead of converging on a registered state library.

## Background Result

BACKTOON did not improve the target background grammar. With Animagine it introduced
rough pastel texture; with Illustrious it collapsed to an under-detailed gray diagram.
Plain Animagine created the sharpest flat plate, but room semantics and prop placement
were wrong. Backgrounds therefore need separate layout or lineart guides and approved
clean plates, not character prompting and not BACKTOON.

## Decision

Do not generate the 28-state pack with any tested public recipe. Keep Animagine plus
iroiro only as the provisional base for internal training and workflow plumbing.

The next quality path is:

1. Author 24-40 original, rights-cleared examples that define our own flat comedy
   style across characters, expressions, poses, and simple backgrounds.
2. Train an internal SDXL style LoRA on that set.
3. For recurring characters, train a separate identity LoRA from approved turnaround
   and expression references.
4. Use OpenPose or lineart structure guides, but generate only neutral, action, and
   extreme reaction until identity passes.
5. Build backgrounds from scene layout guides and reusable clean plates, then composite
   characters separately.

The benchmark remains a timing and production-quality reference only. Its frames must
not be used as the internal LoRA training dataset.

## Sources

- Illustrious XL v2.0: https://huggingface.co/OnomaAIResearch/Illustrious-XL-v2.0
- Xinsir OpenPose SDXL: https://huggingface.co/xinsir/controlnet-openpose-sdxl-1.0
- ComfyUI IPAdapter Plus: https://github.com/cubiq/ComfyUI_IPAdapter_plus
- Emotional Flat SDXL LoRA: https://huggingface.co/ramel2/emotional-flat-illustration-sdxl-lora