A character LoRA for MiniMax H3, a ~33B-parameter omni-modal model that
generates video and audio jointly. Trained on the Ref2VA variant and applied at
inference to its GGUF Q3_K_M build. Trigger token: masafy_character.
The subject is an original red-panda character created by the author. The full
technical report (JA/EN), the training dataset and all measured data are on GitHub:
Training this adapter failed three times before it succeeded, and the reason was
not the one everybody expects. Peak VRAM during training measured 3,959 MiB —
2.8% of the card in use. What actually blocked the run was system RAM: the
41.2 GB model set (20.0 GB transformer + 15.7 GB Qwen3-VL 32B text encoder +
5.5 GB VAEs) is staged in host memory before it reaches the GPU, so hosts with
50 GB of RAM were killed by the OOM reaper while a host with 188 GB completed.
Rent by system RAM, not by VRAM. Advertised RAM tracks neither price nor VRAM.
Checkpoints
File
Steps
Notes
masafy_h3_ref2va_lora_v1_000000300.safetensors
300
style already largely acquired
masafy_h3_ref2va_lora_v1_000000400.safetensors
400
masafy_h3_ref2va_lora_v1_000000500.safetensors
500
recommended
masafy_h3_ref2va_lora_v1_000000600.safetensors
600
masafy_h3_ref2va_lora_v1_000000700.safetensors
700
mild degradation
masafy_h3_ref2va_lora_v1.safetensors
800
final; no better than 500
Recommended: checkpoint 500 at strength 0.8. Cross-validation over strength
(0.0–1.2) and training length (300–800) put the optimum there. Strength trades
style fidelity against structural accuracy: raising it moves the style toward the
target while degrading hands, and at 1.2 a heavy dark contour appears around the
character — the die-cut border of the sticker artwork it was trained on.
800 steps at strength 1.0, the setting used initially, was excessive in both.
Usage (ComfyUI)
Place a checkpoint in ComfyUI/models/loras/ and insert LoraLoaderModelOnly
after UnetLoaderGGUF:
Re-point both BasicGuider and BasicScheduler. Patching only one leaves the
guider and the scheduler looking at different models, and it fails quietly.
Prompt with the trigger followed by the appearance:
masafy_character, a cute anthropomorphic red panda wearing brown aviator goggles
and a pink patterned scarf, waving at the camera with a big smile, standing in a
sunny park, smooth character animation, clean bright colors
The trigger alone is not sufficient. With 12 training images all captioned
with the same appearance clause, the caption text carries real weight; omitting
the appearance description does not produce a usable result.
What the adapter contributes
Under reference-image (R2V) conditioning the reference already pins the character:
without the adapter the goggles, the scarf and the pose all render correctly and
without flicker. What changes is the style — flat cel illustration with the
adapter, photoreal fur without it. The return on this adapter is rendering the
character in the author's own style, not rendering the character at all.
The adapter also reshapes backgrounds, because captions described whole images.
Requirements
Inference VRAM
9.3 GB (fits a 12 GB consumer GPU)
Generation time
~13 min per 5 s clip at 864×480 on an RTX 3060
Base model
MiniMax H3 Ref2VA — not redistributed here
Quantization
trained on pruned INT8, applies cleanly to GGUF Q3_K_M (208/208 modules, zero key mismatches)
License and attribution
The adapter is a derivative work of MiniMax H3 and its use is subject to the
MiniMax H3 Community License Agreement
(text).
Commercial use is permitted and requires attribution to "MiniMax H3". The
license carries a territorial clause; consult the official terms for your
jurisdiction.
The character "Masafy" is an original character created by the author. The design
remains the author's work — please credit the author if you use it.
Citation
bibtex
1@techreport{suzuki2026h3lora,
2 author = {Suzuki, Masato},
3 title = {Character LoRA Fine-tuning of a 33B Omni-modal Video Model:
4 System RAM, Not VRAM, Was the Binding Constraint},
5 year = {2026},
6 url = {https://github.com/masafykun/minimax-h3-masafy-lora}
7}
