biomorphrx β Krea Realtime 14B LoRA
Biomorphic / reaction-diffusion video LoRA trained on Wan 2.1 T2V 14B for the Daydream Scope Krea Realtime pipeline.
Training complete β final checkpoint: biomorphrx_krea_14B_2k_fnl.safetensors (step 2000, ~9 hr on A100 SXM 80GB).
Trigger
biomorphrx
Files
| File | Step | Notes |
|---|---|---|
biomorphrx_krea_14B_2k_fnl.safetensors |
2000 | Final β recommended |
biomorphrx_krea_14B_1750.safetensors |
1750 | Alt |
biomorphrx_krea_14B_1500.safetensors |
1500 | Alt |
biomorphrx_krea_14B_1250.safetensors |
1250 | Early |
biomorphrx_krea_14B_1000.safetensors |
1000 | Early |
config.yaml |
β | ai-toolkit training config |
samples/ |
0β2000 | webp animations every 250 steps Γ 4 prompts (36 total) |
Base model
Wan-AI/Wan2.1-T2V-14B-Diffusers β the 14B base that Daydream Scope's Krea Realtime pipeline distills from. Not compatible with LongLive (1.3B) β see biomorphrx_longlive_1-3B for the LongLive variant trained on the same dataset.
Training summary
Same dataset, captions, trigger word, and prompts as the LongLive 1.3B LoRA. Only base model and native spec differ. Direct A/B of the two Scope pipelines.
- Dataset: 53 images + 40 videos (720Γ480, 16 fps, 41 frames β Wan VAE 4k+1 rule)
- Captions: 100% unique, deterministic combinatorial generation (
FORM Γ COLOR Γ BACKGROUND Γ LIGHTING), trigger word first - Steps: 2000, LoRA rank 32 / alpha 32
- Optimizer: adamw8bit, lr 1e-4, EMA 0.99, gradient accumulation 2
- Quantization: qfloat8 (transformer + text encoder)
- Hardware: RunPod A100 SXM 80GB, ~9 hr, ~$13
- Trainer: ostris/ai-toolkit
Usage in Daydream Scope
- Native resolution: 720Γ480
- LoRA Scale: start at 1.0, bump toward 1.2 if underexpressive. 1.5+ tends to overdrive.
- Mode: runtime/dynamic preferred on first test (permanent_merge has historical quirks with some LoRA formats).
- Prompt format: trigger first, full sentence beats trigger-alone.
Sample prompts
biomorphrx iridescent macro bloom with cyan and magenta hues on black background, shallow depth of field
biomorphrx Belousov-Zhabotinsky reaction-diffusion pattern, teal and orange oscillations, slow swirling
biomorphrx organic slime mold growth spreading over dark substrate, time-lapse
biomorphrx radial cluster of cream petals with blue rim light on black, slow drift
See the samples/ folder for step-by-step evolution of all four prompts from step 0 (base model, no LoRA) through step 2000 (final).
- Downloads last month
- 8
Model tree for diegochavez/biomorphrx_krea_14B
Base model
Wan-AI/Wan2.1-T2V-14B-Diffusers