akhaliq/MiniMax-H3-Claymation-Style-LoRA

MiniMax H3 — Claymation Style LoRA (v1)

A style adapter for MiniMax H3 (FL2VA text-to-video) that turns generated video into a hand-sculpted plasticine claymation / stop-motion look: fingerprinted clay texture, sculpted figures, soft studio lighting.

• Trigger word: claymation (lead your prompt with it, e.g. "claymation of a dog chasing a ball across a park, plasticine clay figures, hand-sculpted clay texture, stop-motion animation style")
• 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: 322 captioned 512×512 claymation stills, re-captioned with the claymation trigger and scrubbed of source-specific wording (dataset, derived from Norod78/ClaymationChristmas-blip-captions-1024)
• 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 1.16 s/step on a single A100 80 GB — ~39 min of training, 1h04m total wall-clock including setup and in-training sampling (≈ $2.70)

Checkpoints

Checkpoints saved every 250 steps are in checkpoints/ (steps 250–2000, 8 files, named full_claymation_h3_v1_*_<step>.safetensors). Overbaking is the known failure mode for style LoRAs — pick by the previews in samples/, not by step count. Start with step 1500 and step 2000.

Known limitations

• Trained on stills: the clay look transfers, but stop-motion cadence (stepped 12fps-style motion) is not learned — motion timing stays base-H3. A video-clip v2 may add that.
• Source data is one claymation special; captions were scrubbed, but occasional winter-scene leakage is possible.

Samples

samples/claymation_h3_v1/samples/ holds in-training previews every 250 steps × 4 generic prompts (dog in park / chef cooking / race car / chess players) — no Christmas scenes, which is the leak test.