Skip to content

MODELS.md

Latest commit

 

History

History
37 lines (29 loc) · 1.92 KB

MODELS.md

File metadata and controls

37 lines (29 loc) · 1.92 KB
Raw

Models that work well

Difforum takes a plain MODEL / VAE / CONDITIONING, so any image model ComfyUI loads can drive the Feedback Sampler. These are good starting points. Check each model's page for current versions and licences.

Feedback Sampler (the Deforum look)

Goal Model Settings
Quality SDXL finetunes (e.g. Juggernaut XL, RealVis XL) 20-28 steps, dpmpp_2m karras, cfg 5-7, energy 0.4-0.6
Speed SDXL + DMD2 4-step LoRA, or Lightning 4-8 steps, lcm / euler, cfg 1-1.5, cadence 2
Flux Flux.1 dev / schnell (fp16 / GGUF on Mac) 8-20 steps, cfg 1, use FluxGuidance on the positive prompt
Stylised / light SD1.5 finetunes (DreamShaper, etc.) 20 steps, cfg 7. On some builds, launch with --force-fp32 if frames come out black
Live SDXL-Turbo, SD-Turbo, LCM LoRA 1-2 steps, cfg ~1, 512 px

A depth model for 3D parallax: Depth Anything 3, built into ComfyUI (Load Depth Anything 3 → Run → Render, depth_anything_3_mono_large.safetensors in models/geometry_estimation). Its v2_style render is near = white, Difforum's convention. Depth Anything V2 (Kijai's pack) still works on any depth input.

Video models (via the bridges)

Model Bridge Notes
LTX-2 / 2.5 LTX Guides Setup target LTX. Runs on Apple Silicon with bf16 / GGUF (no fp8 on MPS).
MiniMax H3 H3 Shot 24 fps, 17k+5 frames, up to 20 s per generation.
Wan 2.x VACE Guide Frames → VACE control video + masks Setup target Wan (4k+1, 16 px).

Apple Silicon

  • fp8 checkpoints do not run on MPS. Use fp16 / bf16 or GGUF.
  • Launch with --use-pytorch-cross-attention, and PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 to unlock unified memory.
  • Few-step models are the real speed lever. Setup's max_megapixels keeps long renders inside memory (about 0.35 MP on 16 GB, 0.65 on 24 GB, 1.0 on 32 GB).