Difforum · Model Profile

The hardest question in modern video generation isn't "what prompt" - it's "what model, what quant, what resolution, and what steps can my GPU actually finish?" Difforum Model Profile is the pack's answer: it resolves a full render profile for your hardware so the same workflow runs from a 12GB card up to 32GB+ without you hand-tuning a dozen knobs.

You give it a family (wan22, sd15_animatediff, ltxv, or sdxl), a quality (fast, balanced, quality), and let auto_detect_vram read your actual VRAM (or type it in vram_gb yourself). It returns a DIFFORUM_PROFILE bundle plus summary, width, height, and segment_frames as plain numbers - the last three you can wire straight into your graph's sizing.

The reasoning it does for you

The profiles are conservative tables keyed to VRAM, and they're genuinely sensible. On a 12GB card with Wan 2.2, it picks the 5B GGUF at Q5_K_M, 640×640, ~49 frames per segment, sequential offload, VAE tiling - "fit and run" values, not glory values. On 32GB it moves up to Wan 2.2 14B fp8 at 720p with no offload. The SDXL path is the Classic+ feedback loop (one frame per segment, per-frame img2img), the SD1.5 path is AnimateDiff with an AnimateLCM motion module at 4–8 steps, and LTX is the fast preview workhorse. The quality dial shifts steps and cfg: fast is 4 steps with a Lightning LoRA at cfg 1.0; quality is 20 steps, cfg 4.0, no distill.

Two details that show it's not a toy. The attention input (sdpa vs sage) lets you opt into SageAttention on CUDA. And on Apple Silicon it detects MPS and swaps recommendations automatically - notably, fp8 checkpoints don't work on MPS, so the profile falls back to fp16/GGUF and recommends the few-step models (SD-Turbo/LCM/Lightning) as the real speed lever, since fp16 isn't faster than fp32 on M-series ALUs.

Install and gotchas

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/chillithebillis/Difforum.git difforum

Restart ComfyUI ([Difforum] loaded N nodes; Manager has it too). Deps: numpy.

The honest caveat: the numbers are conservative defaults meant to fit and run, not to max out quality - the pack's own source says so. Treat the profile as a sane starting point, not gospel: if it picks 640×640 but you know 832×480 renders fine for your scene, override it. And note the outputs are recommendations, not downloads - Model Catalog is the sibling that tells you where to get the exact files the profile expects.