RulzTrainer ComfyUI

ComfyUI

Using your LoRA

Copy the trained .safetensors into ComfyUI/models/loras/ and add a standard Load LoRA node after Load Diffusion Model. MOSS-Image, Krea 2, FIBO 2.0, Qwen-Image 2.1, FLUX 3 Klein, Ideogram 4 and MiniMax H3 are supported by current ComfyUI builds, so no plugin is needed. LoKr files (lokr_w1 / lokr_w2) load too; diffusers has no LoKr loader, so a FIBO 2.0 LoKr loads only back into RulzTrainer.

• Krea 2: train on Raw, use on Turbo (8 steps, no guidance) or on Raw.
• FIBO 2.0: train on dev, use on Edit Turbo (about 4 steps).
• Qwen-Image 2.1: use on the base model (25 steps, CFG 1) or Turbo / the Turbo LoRA (8 steps, CFG 1). The file uses diffusers names; ComfyUI maps the feed-forward's gate and up halves onto its fused layer. The training adapter is not in your file and is not needed to use the LoRA.
• FLUX 3 Klein: use on Base (50 steps, CFG 3.5) or on the distilled Klein (16 steps, CFG 1). The file uses BFL's own names with alpha folded in, which ComfyUI and diffusers both read.
• Ideogram 4: use LoraLoaderModelOnly on the conditional model (ideogram4_*, not ideogram4_unconditional_*). The file keeps the checkpoint's names (fused attention.qkv), as ai-toolkit's Ideogram LoRAs do.
• MiniMax H3: a standard Load LoRA on the FL2VA model; the file uses the lora_A / lora_B layout (alpha = rank) that ComfyUI, musubi-tuner and diffusers read. It was trained on stills; how well it carries into motion is not known yet.
• LoRA Isolation: prompt with the joined trigger (gweltman, not gwelt, a man).
• A fine-tune's extracted LoRA is a normal LoRA at the rank you chose.

Likeness scoring nodes

The repo ships a ComfyUI node pack, comfyui/rulztrainer_likeness, that runs the trainer's own face and body scoring on any images, so you can score a finished LoRA's outputs the same way the probes did.

Install: link the folder into ComfyUI's custom_nodes and restart ComfyUI.

ln -s /path/to/RulzTrainer/comfyui/rulztrainer_likeness ~/ComfyUI/custom_nodes/rulztrainer_likeness

On Windows: mklink /J <ComfyUI>\custom_nodes\rulztrainer_likeness <RulzTrainer>\comfyui\rulztrainer_likeness. ComfyUI's Python needs insightface, onnxruntime, mediapipe, transformers and openpyxl; mediapipe 1.1 or newer is recommended. Then load workflows/RulzTrainer Likeness Scoring.json from that folder.

Three nodes:

Node	What it does
Likeness References	The person to score against: shared references, or separate face and body references; files, folders or connected images. gold_face also loads the held-out FaceNet face recognizer (on by default).
Likeness Score	Scores one image, a batch or folders. Gives overall, face, gold face and body averages and every image's own scores, on the normalized scale.
Likeness Export	A toggle that writes an .xlsx (Summary, Images and References sheets) or CSV files to ComfyUI's output folder.

The scale is the trainer's, including the stretched, curved face scale (see Character LoRAs). Optional base-model images set the 0 of the scale, like the step-0 probe does, and a trainer job folder gives that run's exact scale. Paths are as the ComfyUI process sees them (a ComfyUI in WSL reads C:\ai as /mnt/c/ai).

Gold face. The trainer's probes can score each face a second time with FaceNet (VGGFace2), which nothing ever trains on (face_gold). A run that trains against ArcFace (the reward stage, the identity loss, identity alignment) can raise the ArcFace face score without the person looking more alike; the gold face shows when that happens. The nodes score it too, on its own scale shaped like the face's (the FaceNet weights download into the checkout's models/ the first time). overall_face on the Score node picks the face in the overall score: gold as the trainer does when its gold scorer is on, arcface as a run without it, or auto, which follows the trainer job folder when you give one and uses ArcFace otherwise. The overall with ArcFace's face is always in the details and the spreadsheet.