In short
On 1 November 2026 LTX-2.5 was still the most downloaded open image-to-video model (2,441,077 downloads in 30 days, up 22.5% in two weeks) and Qwen-Image 2.1 Turbo nearly doubled to 39,718. Derivatives outrank originals, and user threads document the fixes.
Two weeks after the October trend report, the open model that people download most for an image-to-video character clip is still LTX-2.5, and it grew from 1,993,003 to 2,441,077 downloads in the last 30 days (+22.5%). The number that moved most is smaller: the Turbo build of Qwen-Image 2.1 nearly doubled, from 20,703 to 39,718.
This is a re-measurement, not a new list. We kept the same six Hugging Face repositories as on 18 October, read their counters again on 1 November 2026, and added the threads where users reported what broke. For anyone running a recurring character, the useful part is the second half: which downloads turn into problems you can avoid.
Trend 1: video still leads, and the lead is growing
LTX-2.5 added about 448,000 downloads to its 30-day count in 14 days. Lightx2v's 8-step MiniMax H3 Turbo grew from 1,497,942 to 1,673,048 (+11.7%). Both are image-to-video models, the kind a character feed uses to animate an approved still, and both are far ahead of every image model on the list.
That gap says something about what creators install, not about quality: a video model is a bigger set of files, so each new user fetches more of them. Treat the ranking as a map of interest. Our guide to open video models for vertical reels covers which of them fit a phone-shaped clip.
Source: Hugging Face model pages, read 2026-10-18 and 2026-11-01
| Model | 18 Oct 2026 | 1 Nov 2026 | Change |
|---|---|---|---|
| LTX-2.5 | 1,993,003 | 2,441,077 | +22.5% |
| MiniMax H3 Turbo (lightx2v) | 1,497,942 | 1,673,048 | +11.7% |
| Qwen-Image 2.1 | 171,381 | 212,670 | +24.1% |
| Krea 2 Raw | 90,298 | 88,816 | -1.6% |
| Krea 2 Turbo | 87,941 | 83,285 | -5.3% |
| Qwen-Image 2.1 Turbo | 20,703 | 39,718 | +91.8% |
Trend 2: the Turbo twin is where the new users land
Qwen-Image 2.1 Turbo nearly doubled in two weeks while the base model grew by a quarter. The same pattern showed up in video: the most downloaded MiniMax H3 file on the board is a Turbo LoRA or a pruned, distilled build, not the full model. Our explainer on distilled versus base checkpoints says when that trade is safe for a face.
Krea 2 went the other way. Raw and Turbo both slipped (-1.6% and -5.3%), which is inside what a rolling window can do on its own. We do not read that as a decline, only as a model that stopped being new.
Trend 3: the biggest download is not the original
On the text-to-image trending list the repository with the most downloads last month is a third-party quantized repack of Qwen-Image 2.1, at 2,558,497 downloads. The official Qwen-Image-2.1 repository has 212,670. That is about twelve times as many fetches for a derivative as for the model it is built from.
Why it matters: a download chart ranks files, and one popular repack can outweigh a whole official release. Before copying what is "trending", open the repository, check who published it and read the license of the parent model, because a derivative inherits the parent's terms. We link only the official repositories in this report.
Trend 4: what Civitai users download is a different crowd
Civitai's most downloaded list for the week is dominated by style LoRAs and checkpoints, with base models now listed as Krea 2, MOSS Image, Anima and MiniMax H3. The model families from the video and image charts above show up there as tags, which is a quick way to see which new bases people have started to train on.
| Model page | Type | Listed base | Downloads all time |
|---|---|---|---|
| Add Micro Details - Concept | LoRA | Krea 2 | 110,721 |
| MiaoMiao RealSkin | Checkpoint | MOSS Image | 65,954 |
| Hyperdetailed Colored Pencil | LoRA | MiniMax H3 | 44,841 |
| Anima Workflows | Workflows | Anima | 39,954 |
| Niji Sweet Spot | LoRA | MOSS Image | 32,979 |
| Detailifier | LoRA | MOSS Image | 29,201 |
| Krea 2 Pro Grade workflow | Workflows | Krea 2 | 23,926 |
Two pages are on both readings. "Add Micro Details" went from 107,654 to 110,721 (+2.8%) and the Krea 2 Pro Grade workflow from 21,942 to 23,926 (+9.0%). A packaged workflow gaining nine percent in two weeks, on an all-time counter, is a steady pull. See how to keep a character consistent in ComfyUI for how to choose between a workflow, a LoRA and reference images.
Trend 5: the problems are in the details, and users document them
The fastest way to learn what a download will cost you is the discussion tab. In the last two weeks the threads on the most downloaded repositories describe four specific problems, each with a workaround from another user.
What creators are saying
A video editor found banding in saturated areas after a VAE round trip on LTX-2.5. A reply says to use the full-precision VAE inside the checkpoint, because the separate bf16 file has already lost the precision.
A user counted that naming someone "out of frame" in an LTX-2.5 image-to-video prompt brought that person into the shot in 6 of 10 prompts. Their workaround: describe gaze direction and visible light instead.
ComplaintLTX-2.5 discussion 92, Hugging Face
On a 12 GB card, MiniMax H3 reference-to-video clips looked slow and weightless. Replies suggest switching to first-and-last-frame mode with the keyframe, a motion LoRA, and more energetic verbs.
A user reports that editing with the same seed as the reference image gives distorted output on Qwen-Image 2.1, because the edit path reuses a noise pattern correlated with the reference. Changing the second seed fixed it.
Someone asks which license covers the new Ref2V Turbo weights, because the model card lists none and they want to use them in a small commercial video project.
A user says Krea 2 Turbo is good but weak with anime styles, and compares it with the Anima base model.
For a character feed the Qwen seed report is the one to remember: if you generate a still and then edit it, use a different seed for the edit. It costs nothing and the thread says it removes a failure that looks like a prompt problem. The LTX off-frame result comes from one user's clips, so treat 6 of 10 as an anecdote with a useful lesson, not a measured rate.
Trend 6: licenses lag the files
The Ref2V Turbo question is a small version of a larger pattern: a LoRA or repack appears with no license on its card, and the person who wants to publish with it has to ask. Another LTX-2.5 thread asks about licensing its outputs as stock footage. Neither thread had an authoritative answer when we read it.
Cheat-sheet: this fortnight at a glance (1 Nov 2026)
- Most downloaded image-to-video
- LTX-2.5, 2,441,077 downloads last month (Hugging Face)
- Fastest mover on our list
- Qwen-Image 2.1 Turbo, +91.8% in 14 days
- Biggest file is not the official one
- A third-party Qwen-Image 2.1 repack, 2,558,497 vs 212,670 for the original
- Civitai base tags in the top list
- Krea 2, MOSS Image, Anima, MiniMax H3
- Edit after generate (Qwen)
- Use a different seed for the edit
- LTX-2.5 banding
- Use the full-precision VAE from the checkpoint
- MiniMax H3 slow motion
- Try first-and-last-frame mode, a motion LoRA, energetic verbs
What it means for a character feed
Pick by the face, not by the chart. A model that gains downloads can still change your character's look, so run a trial the same way every time: same approved references, four draft seeds on the new checkpoint, then finish on the one you trust. The consistent character guide and the LoRA training guide describe the steps.
Musetta is a studio that runs on your own ComfyUI and keeps each character's approved references, so trying a new model does not change who your character is. We make it, so weigh that claim accordingly. The screenshot below is the studio staged with demo data.

Gaps and caveats
- Two readings are not a trend line. Two weeks of a rolling window is a direction, not a forecast.
- We could not read Reddit from this environment, so voices come from Hugging Face threads only.
- Download counts include automated and repeat fetches, and repacks split the audience.
- Civitai counts are all-time, Hugging Face counts are 30-day; the two lists are never added together.
Our verdict
The signal this fortnight is not a new champion. It is that the Turbo twins keep pulling new users, that the biggest downloads are often derivatives, and that the problems with each model are documented by users within days. Read the discussion tab for the model you are about to install; it is faster than any benchmark. The mid-November re-reading adds MiniMax H3 Image's first week. If you want the weekly picture, the October report and yesterday's ComfyUI 0.43 news are the baseline. The newest ComfyUI build at the time of reading was v0.43.1 (28 October), per the ComfyUI changelog, so update before you test a new model. For the repack-versus-official gap measured across twelve model families, see the repack data study.
Questions
- What is the most downloaded open model for image to video right now?
- In the 1 November 2026 reading, Lightricks LTX-2.5 led with 2,441,077 downloads in the last 30 days on Hugging Face, up 22.5% from 1,993,003 on 18 October. Counts measure file downloads, not users.
- Which model grew fastest in the last two weeks?
- Among the six repositories we tracked, Qwen-Image 2.1 Turbo, from 20,703 to 39,718 downloads in 30 days (+91.8%). Qwen-Image 2.1 grew 24.1%.
- Why is a repack more downloaded than the official model?
- Hugging Face counts file fetches per repository. One popular quantized repack can be fetched far more often than the official weights, so a trending list ranks files, not model quality or origin.
- Do Hugging Face and Civitai download numbers match?
- No. Hugging Face shows a rolling 30-day count per repository and Civitai shows all-time downloads per model page, so the numbers cannot be added or compared directly.
- How do I avoid distorted edits in Qwen-Image 2.1?
- One user reports that using the same seed for the reference image and the edit gives bad output because the edit path reuses a correlated noise pattern, and that a different seed for the edit fixed it. It is a user report, not an official statement.
Sources
- 01Hugging Face trending models · Hugging Face, 2026-11-01
- 02LTX-2.5 model card · Lightricks, Hugging Face, 2026-11-01
- 03Minimax-h3-Turbo model card · LightX2V, Hugging Face, 2026-11-01
- 04Qwen-Image-2.1 model card · Qwen, Hugging Face, 2026-11-01
- 05Krea-2-Turbo model card · Krea, Hugging Face, 2026-11-01
- 06MOSS-Image-SFT model card · OpenMOSS, Hugging Face, 2026-11-01
- 07LTX-2.5 discussion 89 · Hugging Face community, 2026-10-19
- 08LTX-2.5 discussion 92 · Hugging Face community, 2026-10-21
- 09Minimax-h3-Turbo discussion 58 · Hugging Face community, 2026-10-20
- 10Minimax-h3-Turbo discussion 60 · Hugging Face community, 2026-10-28
- 11Qwen-Image-2.1 discussion 61 · Hugging Face community, 2026-10-13
- 12Krea-2-Turbo discussion 20 · Hugging Face community, 2026-08-22
- 13ComfyUI changelog, v0.43.1 of 28 October 2026 · Comfy, 2026-11-01
- 14Civitai most downloaded models · Civitai, 2026-11-01
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.