akhaliq/MiniMax-H3-Anime-2D-Style-LoRA

MiniMax H3 — 2D Anime Style LoRA (v1)

A flat 2D anime cel-shading style adapter for MiniMax H3 (FL2VA text-to-video): clean line art, flat cel shading, vibrant anime color work.

• Trigger word: 2danime (lead your prompt with it, e.g. "2danime style anime illustration of a girl walking through a city street at sunset, flat 2D anime cel shading, clean line art, vibrant colors")
• Recommended strength: 0.8–1.0
• Works with: FL2VA text-to-video; attach as a model-only LoRA in ComfyUI on the minimax_h3_fl2va_pruned_int8_convrot weights. Do not merge into the quantized checkpoint — load it as a live adapter.

Training

• Data: 420 anime wallpaper stills sampled across the full history of puruchinera/anime_wallpapers (spread across date folders, 512px short side, filtered for aspect/size), captioned with the 2danime trigger + style descriptors (prep'd dataset)
• Method: ai-toolkit minimax_h3 extension (FL2VA partition, pruned int8-ConvRot DiT), rank 16, single-latent-frame image training at 512px, adamw8bit, lr 1e-4, latents + text embeddings cached to disk, contrastive guidance loss (target 3.5) plus the frozen ostris training adapter, ignore_if_contains: ["adaln_proj"], audio not trained
• Observed run: 2,000 steps at ~1.2 s/step on a single A100 80 GB — 1h08m total wall-clock including setup and in-training sampling (≈ $2.85)

Checkpoints

Checkpoints saved every 250 steps are in checkpoints/ (steps 250–1750 + final 2000). Pick by the previews in samples/, not by step count.

Known limitations

• Trained on wallpapers: scenery and illustration style transfer strongly; character-focused shots (faces, close-ups) may be weaker since wallpapers skew toward wide scenes.
• The source pool is community-curated wallpaper art — occasional non-2D renders (3D-style wallpaper art) may have slipped into training; captions describe the flat 2D target style.
• The look transfers; anime motion (smears, limited animation timing) is not learned — motion stays base-H3.

Samples

samples/anime2d_h3_v1/samples/ holds in-training previews every 250 steps × 4 prompts (schoolgirl street / knight cliff / cat windowsill / fighter jet).