Introduction

Hours-old on Hugging Face: KBlueLeaf/MOSS-Image-heal (created 2026-11-04 ~05:39 UTC) is a full transformer finetune of OpenMOSS-Team/MOSS-Image-Base aimed at one pain point — the base model's residual stream peaks near the FP16 cliff. Heal pulls the largest stream value from 58,411 down to 3,083, cuts stream RMS about 7.5×, and still tracks the base closely (flow loss 0.4289 vs 0.4280, velocity cosine 0.9983 on the card's probe set). Apache-2.0 weights, Diffusers swap, and ready ComfyUI single-file checkpoints in the same card.

What "heal" actually changes

Only the transformer is trained. Text encoder, tokenizer, VAE, scheduler, and pipeline stay the official MOSS-Image-Base pieces. The recipe is stream healing in the KohakuMoss sense: flow matching plus distillation toward a frozen base, plus losses that punish concentrated stream writes and high stream RMS, with a decoupled weight-outlier decay so no element sits above 64× its tensor RMS (base had 349 elements above that bar; heal has 0).

Training snapshot from the model card:

  • 50,000 steps × 64 samples (~3.2M samples), ~200 GPU-hours on B300
  • ~1M quality-ranked Danbooru posts with Graph-Based Captioning (GBC) graphs
  • Prompt formats mix tags, natural language, and gbc.soft layout graphs
  • Text refiner stays frozen; everything else in the transformer moves

Precision check on the card: the FP16 Comfy file peaks around 3,278 in the stream on the probe rows — roughly 20× headroom under the FP16 max — while base peaks near 61,696 on the same rows. That is the practical win for --fp16-unet graphs and INT8 ConvRot packs that used to flirt with overflow.

Where to run it today

ComfyUI (easiest maker path). Download one of the files under comfyui/ into models/diffusion_models/:

  • moss_image_heal_bf16.safetensors
  • moss_image_heal_fp16.safetensors (pair with ComfyUI --fp16-unet)
  • moss_image_heal_int8_convrot.safetensors (Comfy-Org INT8 ConvRot layout)

Keep the usual Comfy-Org Qwen3.5-9B text encoder and FLUX.2 small VAE you already use for MOSS-Image. Load with Load Diffusion Model in the official MOSS-Image workflow — no new custom node pack required.

Diffusers. Load the official MOSS-Image-Base pipeline, then swap in heal's transformer:

import importlib
import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "OpenMOSS-Team/MOSS-Image-Base",
    custom_pipeline="pipeline_moss_image",
    trust_remote_code=True,
    transformer=None,
    torch_dtype=torch.bfloat16,
)
pipe.scheduler = type(pipe.scheduler).from_config(pipe.scheduler.config)
MOSSImageTransformer = importlib.import_module(type(pipe).__module__).MOSSImageTransformer
pipe.transformer = MOSSImageTransformer.from_pretrained(
    "KBlueLeaf/MOSS-Image-heal",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

This is open weights (Apache-2.0, same as official MOSS-Image). Local Comfy / Diffusers today — not a closed API partner node.

Why it matters

MOSS-Image already has native Comfy support and a growing LoRA scene. Heal is the reliability patch for people who want FP16 / INT8 paths without the base stream slamming the float16 ceiling, while keeping samples almost interchangeable with base (the card shows side-by-side caption and GBC grids). If your MOSS graphs crash, wash out, or refuse to stay in FP16 under load, this is the checkpoint to A/B first.

Caveats

  • Brand-new card (★1 at scout time) — treat as day-0 community finetune, not a vendor release
  • Still needs the official MOSS text encoder + FLUX.2 VAE stack; heal does not replace those files
  • License is Apache-2.0 on the heal weights; confirm any other stack pieces you mix in
  • Not a Gen-hosted model on ArtRealmAI today — run local / your own host

Original Source

Conclusion

Drop moss_image_heal_bf16 (or the FP16 / INT8 ConvRot twin) into diffusion_models/, keep your MOSS encoder + small FLUX.2 VAE, and A/B against MOSS-Image-Base on the same seed. If the stream used to flirt with FP16 overflow, heal is the quiet fix.

—Aurelia ♡