In short
Across twelve model families, Comfy-Org's ComfyUI repack beats the official Hugging Face repos on 30-day downloads in eleven. MOSS Image leads at 154 to 1, then Wan 2.2 at 118x and Qwen-Image 2.1 at 41x. LTX-2.3 is the exception at 0.28. Download counts compare traffic, not users or quality.
If you run a character on ComfyUI, the repo you download from decides which file you get, which folder it goes in and which ComfyUI version it needs. On Hugging Face the choice is already made for most people: Comfy-Org's MOSS-Image repack had 4,215,132 downloads in the last 30 days, the two official MOSS repos together had 27,414. That is 154 to 1.
This data study measures that gap for twelve model families, explains what it can and cannot mean, and ends with a rule for picking the repo. Every count below was read from the Hugging Face API on 2 November 2026.
What we measured and how
For each model we took the Comfy-Org repack (the repo whose card says it holds files repackaged for ComfyUI) and the official repos of the lab that trained it. We read the 30-day downloads field of each from huggingface.co/api/models/<repo>, the same figure shown on the repo page. Where a lab publishes several official repos for one family (MOSS SFT and Base, Krea 2 Turbo and Raw, four Wan 2.2 repos), we added them. That makes the official side generous, so the ratios are conservative.
The pairs are the twelve Comfy-Org repos for which we could find the lab's own repo on Hugging Face and that sit in Comfy-Org's top 40 by downloads. We did not pick pairs to prove a point: the one pair that goes the other way (LTX-2.3) is in the table.

The repack wins eleven of twelve
| Model | Comfy-Org repack | Official repos | Official side | Ratio |
|---|---|---|---|---|
| MOSS Image | 4,215,132 | 27,414 | SFT + Base | 153.8x |
| Wan 2.2 | 9,185,113 | 77,764 | four repos | 118.1x |
| Qwen-Image 2.1 | 10,554,957 | 258,167 | base + Turbo | 40.9x |
| Krea 2 | 4,068,766 | 174,388 | Turbo + Raw | 23.3x |
| Z-Image Turbo | 9,441,612 | 747,729 | one repo | 12.6x |
| Wan 2.1 | 4,629,464 | 474,062 | four repos | 9.8x |
| Qwen-Image | 2,585,400 | 324,445 | one repo | 8.0x |
| MiniMax H3 | 30,246,865 | 5,924,727 | one repo | 5.1x |
| FLUX.2 dev | 1,113,492 | 332,388 | one repo | 3.3x |
| LTX-2 | 911,528 | 298,977 | one repo | 3.0x |
| FLUX.1 dev | 1,211,475 | 549,689 | one repo | 2.2x |
| LTX-2.3 | 244,764 | 888,151 | one repo | 0.3x |
Eleven of the twelve repacks out-download the official repos. Four of them by more than 20 to 1: MOSS Image (154x), Wan 2.2 (118x), Qwen-Image 2.1 (41x) and Krea 2 (23x). The Wan 2.2 figure counts four official repos (T2V-A14B, I2V-A14B, S2V-14B and TI2V-5B); Wan-AI publishes others we did not add, so treat it as an upper bound.
The exception is LTX-2.3, where the official Lightricks repo has 888,151 downloads and the Comfy-Org one 244,764, a ratio of 0.28. Its newer sibling LTX-2.5 has no Comfy-Org repo in the top 40 at all: the official Lightricks/LTX-2.5 repo has 2,431,846 downloads and is the biggest video number in the pulse. So the repack is not a law of nature. It is what happens when the official repo is not shaped for ComfyUI.
Why the gap exists
Three differences show up in the repo metadata, none of which needs a theory.
- Layout. The official MOSS, Qwen-Image 2.1, Krea 2 and Z-Image repos carry the
diffuserslibrary tag and hold a pipeline folder tree (MOSS-Image-SFT lists 37 files). The Comfy-Org repos carrydiffusion-single-fileand hold the loader-ready files (MOSS-Image lists 13). - The docs point there. The ComfyUI tutorial for MOSS-Image tells you to download from Comfy-Org/MOSS-Image: two int8 checkpoints, a
qwen3.5_9b_w4a8_gptqtext encoder and a small Flux 2 VAE, with bf16 builds for more VRAM. The templates you open in ComfyUI link the same repo. - Gating is not the reason. FLUX.1-dev, FLUX.2-dev and Krea 2 are gated on the official side (
auto), and their ratios are 2.2x, 3.4x and 23x. MOSS and Z-Image Turbo are not gated and sit at 154x and 12.6x. LTX-2.3 is not gated and goes the other way.
One more factor sits outside our data. The Hub counts requests to a repo's query files, and which file that is depends on the library tag. We did not verify which file each repo counts, so a ratio compares counters, not people. Treat 154 to 1 as 'by far the larger traffic', not as 154 times more users.
Repacks are a growing share of the whole Hub
The pattern is not only a ComfyUI one. A Hugging Face community post by Tardelli Stekel, published 12 October 2026, reports that derivatives (models whose card declares a base model) rose from 18% of Hub downloads in March 2025 to 43% in September 2026, and quantizations from 9% to 28%. We did not rebuild that analysis, so we attribute it to its author. It fits what we see in the ComfyUI corner: the file people actually fetch is increasingly a rebuilt version of the lab's release.
What creators are saying
What creators are saying
On stable ComfyUI 0.41.3 the MOSS CLIP type was missing; users were told stable lacked MOSS support and to use the nightly build, after backing up their install.
A dtype error at the sampler went away when the SD3 shift node was swapped for the AuraFlow or Flux one; staff said the model file was not the cause.
Launching with the sage-attention flag gave desaturated images with artifacts; removing the flag fixed it.
Someone asked where the fp8 quants of Qwen-Image 2.1 are, since the repack lists bf16 and int8 files.
A request for an NVFP4 build of Qwen-Image 2.1.
Look at what these have in common: the questions are about the repack, because that is where people are. Version floors, shift nodes and launch flags decide whether the repack works, and the official repo will not tell you any of it.
Which repo to download from
Download rule of thumb
- You run the model in ComfyUI
- Comfy-Org repack, and check the tutorial for the version floor (MOSS-Image needs ComfyUI 0.42.0 or later)
- You run it in Diffusers or your own code
- The official repo, which carries the pipeline layout
- You need the license text
- The official model card, then compare the repack's license tag (it matched in the ten pairs we compared)
- You want fp8, NVFP4 or GGUF
- Check the repack file list first; ask in its discussions before searching third-party repos
- You judge a model by its download count
- Do not. Use it to find where the files live, then test on your own face
For a character feed the practical point is boring and useful: the files in the Comfy-Org repo are the ones the official templates expect, so a template that loads without red nodes is a better sign of the right download than any count.

Limits of this study
- One reading, 2 November 2026. Counts move daily; the MOSS repos are about three weeks old, so their 30-day figure is nearly their whole life.
- Download counts are request counters on query files, and which file counts differs by library tag. We did not verify per-repo counting files.
- The official side sums the repos we found. Labs can publish others (GGUF, LoRAs, distilled builds) that we did not add.
- We selected pairs from Comfy-Org's top 40 by downloads, so models without a Comfy-Org repack are out of scope.
- Nothing here measures quality, and nothing measures unique people.
Related reading: how big the downloads are, the two-week download re-measure, the MOSS Image review and the fix for missing nodes. If your own install is the bottleneck, renting a GPU is the other route.
Questions
- Should I download from Comfy-Org or the official Hugging Face repo?
- For ComfyUI, the Comfy-Org repack: its files match the loader nodes and the official templates. For Diffusers or your own code, the official repo, which has the pipeline layout. Check the license on the official model card in both cases.
- Does 154x mean 154 times more people use the repack?
- No. Hugging Face counts every GET and HEAD request to a repo's query files as a download, and the query file depends on the library tag. The ratio compares counters, so read it as much larger traffic, not as a headcount.
- Why does LTX-2.3 go the other way?
- Its official Lightricks repo has 888,151 downloads and the Comfy-Org repo 244,764. We did not test why. Both repos are ungated; the official repo has 11 files, so a small, loader-shaped layout may be enough there.
- Are repacks allowed to be redistributed?
- Each repo carries a license tag on its card. In the ten pairs we compared, the repack's tag matched the official one, for example Apache 2.0 for MOSS-Image and 'other' for Qwen-Image 2.1. Read the upstream license text before using outputs commercially.
Sources
- 01Hugging Face Hub docs: Models Download Stats · Hugging Face
- 02Comfy-Org repos on Hugging Face, sorted by downloads (API) · Hugging Face, 2026-11-02
- 03Comfy-Org/MOSS-Image · Comfy-Org, 2026-11-02
- 04OpenMOSS-Team/MOSS-Image-SFT · OpenMOSS, 2026-11-02
- 05MOSS-Image in ComfyUI: Text-to-Image and Layout Workflows · ComfyUI docs
- 06Comfy-Org/Qwen-Image-2.1 · Comfy-Org, 2026-11-02
- 07Model Pulse: daily download history for every model on the Hugging Face Hub · Tardelli Stekel, Hugging Face community blog, 2026-10-12
- 08Model Hubs and Beyond: Analyzing Model Popularity, Performance, and Documentation · arXiv, 2025-03-19
Researched and written by the Musetta editorial desk with AI assistance, from the sources above. Numbers come from those sources; nothing is benchmarked by us unless the article says so. Spotted an error? Tell us on Discord and we fix it.