Global site · runninghub.ai Promoted Sign up via our link — 1,000 RH Coins free Campaign terms on the platform page; we may earn a commission at no extra cost Sign up · 1,000 RH Coins → China site · runninghub.cn Promoted Sign up via our link — 1,000 RH Coins free Campaign terms on the platform page; we may earn a commission at no extra cost Sign up · 1,000 RH Coins → Only models that load directly in ComfyUI, screened by downloads, likes and community reputation (data from HuggingFace). ★ = editor’s pick. Not sure which to get? Run the calculator for your GPU first.
All files used by this site's calculator come from here: bf16 / int8_convrot / pruned / fp8_scaled + TE + VAE + turbo LoRA + ControlNet pack
Unpruned GGUF, Q2-Q8 tiers, loads directly with ComfyUI-GGUF nodes; even 8GB VRAM can give it a try
Pruned + GGUF, Q3_K_M is only 8.9GB (data source for this site's VRAM-saving tier)
Turbo LoRA and Ref2VA delta fused straight into INT8 weights: one less LoRA load, more stable VRAM, ready for 4-8 steps out of the box
8-step distillation, the community's best-reviewed general accelerator; use with the dual-clock sampler
ComfyUI-compatible conversion of larry/lightx2v, with dynamic-rank compressed variants (rank28 is only 0.37GB)
The only source for this site's W4A8 (12.5GB), 1.5x int8 VAE, FastH3 VSA int8, and the PDD comfy conversion
Realistic/cinematic people finetune; README explicitly supports ComfyUI (Load LoRA works as-is, no conversion), trigger word r34l1sm
GGUF build of the Qwen3-VL-32B text encoder, loaded with ComfyUI-GGUF, easing encoder residency issues on low-RAM machines
One of the highest-download community GGUF sources, with Unsloth dynamic quants; loads via ComfyUI-GGUF
Latent-space upscaling/enhancement for generated clips (one of the community 2K routes); safetensors, ComfyUI-loadable
Community hybrid pruned merge: one file for both task modes, loads directly in ComfyUI; among this week’s fastest-growing merges (25.5k downloads)
Multi-format quantization collection (NVFP4 needs 50-series); good for quick comparison
Ref2VA delta fused into pruned INT8; one file covers both modes and saves disk space and load switching
Community W4A8 (4-bit weights + INT8 compute) conversion; requires ComfyUI ≥0.31
Early NVFP4 conversion (Blackwell), ~12GB
Parallel Decoding Distillation, CFG-free; requires the ComfyUI PDD node pack (the nodes can read the official files directly) and cannot stack with other distillation
ComfyUI key conversion of the alibaba-pai PDD weights (for use with the PDD node pack)
ComfyUI-loadable LoRA for the FastVideo 4-step distillation, plus companion nodes that route attention to the VSA sparse path (with the 4-step distillation they give minute-level drafts on 24GB)
Highly upvoted all-round community finetune with an HDR/texture lean; check its usage terms and license
Boosts physical realism — collisions, gravity, stacking — at strength 0.3-0.5 and stacks with turbo; trained on Comfy-Org weights and works with the built-in loader
Stable camera work and push/pull/pan/tilt control, good for strengthening shot language; trained on Comfy-Org weights
Pose/depth/edge control nets (H3-Fun series); Comfy-Org has packed them in native ComfyUI format (model_patches/), the original Diffusers version is not listed here
Community GGUF collection based on Comfy-Org weights — an alternative quant source
Swap the character in a reference video for a given identity (Ref2VA route); ComfyUI-loadable, needs a reference video
360° orbit camera moves in first-last-frame mode; good for product and character orbits
Depth-warp camera re-projection (CrossView-Warp v1); experimental, suits re-camera and re-lighting shots
Community merge of both modes into a single file (fast-growing likes); verify format and license before use
A new distillation route (8 NFE with a tile-score predictor), T2VA preview; experimental, not yet on the arena leaderboard
360° card/product orbit (Ref2VA route) for e-commerce showcases; freshly released (2026-09-28)