diegochavez's picture
Upload folder using huggingface_hub
caf3e56 verified
|
Raw
History Blame Contribute Delete
1.92 kB
metadata
license: apache-2.0
tags:
  - lora
  - wan2.1
  - text-to-video
  - longlive
  - daydream-scope
base_model: Wan-AI/Wan2.1-T2V-1.3B-Diffusers

biomorphrx — LongLive 1.3B LoRA

Biomorphic / reaction-diffusion video LoRA trained on Wan 2.1 T2V 1.3B for the Daydream Scope LongLive pipeline.

Trigger

biomorphrx

Files

File Step Notes
biomorphrx_lnglve_1-3B_2k_fnl.safetensors 2000 Final — recommended
biomorphrx_lnglve_1-3B_1750.safetensors 1750 Alt
biomorphrx_lnglve_1-3B_1500.safetensors 1500 Alt
config.yaml — ai-toolkit training config
samples/ 0–2000 webp animations every 250 steps × 4 prompts

Base model

Wan-AI/Wan2.1-T2V-1.3B-Diffusers — the 1.3B base that Daydream Scope's LongLive pipeline distills from. Not compatible with Krea Realtime (14B) pipelines.

Training

  • Dataset: 53 images + 40 videos (832×480, 16 fps, 77 frames — 4k+1 Wan VAE rule)
  • Captions: 100% unique, combinatorial generation (FORM × COLOR × BACKGROUND × LIGHTING), trigger word first
  • Steps: 2000, LoRA rank 32 / alpha 32
  • Optimizer: adamw8bit, lr 1e-4, EMA 0.99
  • Hardware: RunPod A100 SXM 80GB, ~4 hr
  • Trainer: ostris/ai-toolkit

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

Native resolution: 832×480. 77 frames at 16 fps. LongLive is autoregressive — long output comes from rollout, not training clip length.