In short
On 15 November 2026 MiniMax H3 Image dominated the Hugging Face text-to-image list five days after its weights went public: Comfy-Org's repack logged 1,615,613 downloads, 3.1 times the official repository's 517,948. LTX-2.5 still leads image to video at 2,777,667 (+13.8% in two weeks), Qwen-Image 2.1 grew 28.8% and its Turbo slipped 5.7%.
Five days after its weights went public, MiniMax H3 Image owns the Hugging Face text-to-image board, and the file people fetch is not MiniMax's own. Comfy-Org's ComfyUI repack logged 1,615,613 downloads by 15 November 2026; the official repository logged 517,948. If your character's next set is about to move to a new base model, this is the fortnight that decides which files and which license you will be living with.
This is the third reading of the same counters. We re-read the six repositories from the 1 November report, added the new MiniMax H3 Image family, and read the Hugging Face threads where people who downloaded it reported what happened next.
Trend 1: a new image model reached the video leaders in a week
Until this fortnight, every chart we drew had the same shape: two video models far ahead, image models a long way behind. MiniMax H3 Image broke it. Its four most downloaded repositories (the Comfy-Org repack, the official weights, Abiray's GGUF quants and akhaliq's 2-step pDMD LoRA) add up to 2,269,894 downloads, most of them in the days since MiniMax put the weights online on 10 November.
LTX-2.5, the image-to-video leader, logged 2,777,667 over a full 30 days. The H3 Image family is within 20% of that with a fraction of the time on the board. The ComfyUI 0.45 release that put H3 Image in core, covered in yesterday's news, is the obvious reason: a model that loads from a stock template gets downloaded by everyone who opens the template.
Trend 2: the repack beats the original, again
Comfy-Org's repack has 3.1 times the downloads of MiniMax's official repository. Pruna's ComfyUI build of P-H3-Image shows the same pattern at an extreme: 1,116,266 downloads against 1,952 for Pruna's own diffusers repository, about 572 to one. Our data study on repacks measured 154 to one for other models on 2 November; the gap is not closing.
For a character feed this matters for two reasons. The ComfyUI repacks are what the stock templates point to, so they are what you will actually run. And a repack carries the parent's license: Comfy-Org's and Abiray's H3 Image pages both list the MiniMax-H3-Image community license, the same terms as the official weights.
| Repository | 1 Nov 2026 | 15 Nov 2026 | Change |
|---|---|---|---|
| Lightricks/LTX-2.5 | 2,441,077 | 2,777,667 | +13.8% |
| lightx2v/Minimax-h3-Turbo | 1,673,048 | 1,818,924 | +8.7% |
| Qwen/Qwen-Image-2.1 | 212,670 | 273,962 | +28.8% |
| krea/Krea-2-Raw | 88,816 | 107,299 | +20.8% |
| krea/Krea-2-Turbo | 83,285 | 85,580 | +2.8% |
| Qwen/Qwen-Image-2.1-Turbo | 39,718 | 37,440 | -5.7% |
| OpenMOSS-Team/MOSS-Image-SFT | 16,291 | 23,809 | +46.1% |
| Comfy-Org/MiniMax-H3-Image | new (8 Nov) | 1,615,613 | new |
| PrunaAI/P-H3-Image-Comfy | not tracked | 1,116,266 | new to list |
| MiniMaxAI/MiniMax-H3-Image | not public | 517,948 | new |
| Abiray/MiniMax-H3-Image-GGUF | new (10 Nov) | 111,241 | new |
Trend 3: the Turbo twin cooled, the base models came back
Two weeks ago Qwen-Image 2.1 Turbo nearly doubled and we wrote that the Turbo twins were where new users land. This fortnight reversed it: the Turbo build slipped 5.7% while the base Qwen-Image 2.1 grew 28.8%. Krea 2 Raw, which we had filed under "stopped being new", rose 20.8%, while Krea 2 Turbo barely moved.
Our reading, and it is only a reading: a model that runs at full quality in six steps, as H3 Image does, takes away the main reason to choose a distilled twin. When the new default is already fast, people go back to the base weights to train on and fine-tune. MOSS-Image-SFT growing 46.1% fits the same story. Our explainer on distilled versus base checkpoints covers when each is safe for a face.
Trend 4: Civitai creators are re-basing their LoRAs
On Civitai the list is still style LoRAs and checkpoints, but the base-model tags moved. A MiniMax H3 checkpoint, DaSiWa MiniMax H3, is in the week's top five with 80,450 all-time downloads. Detailifier, a MOSS Image LoRA on 1 November, now lists Qwen 2.1 on a version dated 13 November, and the Krea 2 Pro Grade workflow published a new version the same day.
| Model page | Type | 1 Nov 2026 | 15 Nov 2026 | Change |
|---|---|---|---|---|
| Add Micro Details - Concept | LoRA | 110,721 | 113,670 | +2.7% |
| MiaoMiao RealSkin | Checkpoint, MOSS Image | 65,954 | 68,925 | +4.5% |
| Hyperdetailed Colored Pencil | LoRA, MiniMax H3 | 44,841 | 46,433 | +3.6% |
| Detailifier | LoRA, now Qwen 2.1 | 29,201 | 30,388 | +4.1% |
| Krea 2 Pro Grade workflow | Workflow, Krea 2 | 23,926 | 26,029 | +8.8% |
The packaged workflow grew fastest again, by nearly 9% in two weeks on an all-time counter. People still pay for a graph that someone else has already made work. That is the steady signal under the launch noise.
Trend 5: what the first week of H3 Image looks like from the threads
The discussion tabs on the H3 Image repositories filled within days. Three things stand out for anyone running a recurring character: it runs on a 12 GB card with enough system RAM, it is fast, and it is not yet trustworthy for keeping a face from a reference.

What creators are saying
On a 12 GB RTX 4070 SUPER, the int8 files ran at 4.1 MP in 11-12 s per image warm. Against Qwen-Image 2.1, H3 Image followed prompts better but gave flatter light and simpler skin.
Editing with two reference images of two people, one of the faces came out different from its source.
ComplaintComfy-Org/MiniMax-H3-Image discussion 11, Hugging Face
Someone who builds character datasets found the official ComfyUI node changed appearance and body proportions in edits, and stayed with P-H3-Image, which followed references better.
ComplaintPrunaAI/P-H3-Image-Comfy discussion 12, Hugging Face
A user hit an unknown node pack after downloading gigabytes and called the release rushed. The reply: update to the latest stable ComfyUI and it works out of the box.
Several people asked whether images made locally may be posted on Instagram or X, which are seen in the excluded territories, and whether small commercial projects are allowed.
QuestionMiniMaxAI/MiniMax-H3-Image discussion 3, Hugging Face
A user says the name makes the model hard to research, because search results mix it up with MiniMax's H3 video model.
ComplaintMiniMaxAI/MiniMax-H3-Image discussion 27, Hugging Face
The face reports are the ones to weigh. They come from a handful of users, not a test, but they point the same way: for edits from a reference, P-H3-Image or a character LoRA is still the safer path, and our H3 Image review reached the same caution from the model card's side.
Trend 6: the license questions arrived before the answers
MiniMax opened its own license Q&A thread on launch day. Its first answer confirms the territory scope: the open-weight license does not cover the EU, the UK, South Korea or the US, which need separate authorization. The follow-up questions in that thread (posting outputs to global platforms, running on a cloud GPU whose region you cannot verify) had no official reply when we read it on 15 November.
Cheat-sheet: mid-November 2026 at a glance (read 15 Nov)
- Most downloaded image-to-video
- LTX-2.5, 2,777,667 in 30 days, +13.8% in two weeks
- Fastest new arrival
- Comfy-Org H3 Image repack, 1,615,613 since 8 Nov
- Repack vs official, H3 Image
- 1,615,613 vs 517,948 (3.1x)
- Repack vs official, P-H3-Image
- 1,116,266 vs 1,952 (about 572x)
- Base models up, Turbo down
- Qwen-Image 2.1 +28.8%, Krea 2 Raw +20.8%, Qwen 2.1 Turbo -5.7%
- Runs on 12 GB (one user)
- int8 H3 Image, 11-12 s per 4.1 MP image, about 41 GB system RAM
- Face from a reference
- Users report drift in H3 Image edits; P-H3-Image or a LoRA is safer
- License
- Excludes EU, UK, Korea, USA; authorization above 1M USD revenue
What it means for a character feed
A launch fortnight is the worst time to move a character to a new base, and the best time to test one. Draft on H3 Image if your territory and revenue allow it, keep your approved references fixed, and compare four seeds against your current model before anything is posted. The consistent character guide describes the comparison. If the license rules H3 Image out, our ranking of open image models for a recurring face puts the alternatives in order.
Musetta is a studio that runs on your own ComfyUI and keeps each character's approved references, so a trial on a new model starts from the same face. We make it, so weigh that accordingly. The screenshot below is the studio staged with demo data.

Gaps and caveats
- Three readings are a direction, not a forecast; a 30-day window shifts every day.
- New repositories show their whole life as "last month", so they are not comparable to a full month of an older model.
- Reddit refused this environment, so voices come from Hugging Face threads, each one user's report.
- Civitai counts are all-time and Hugging Face counts are 30-day; we never add them together.
- We did not run H3 Image ourselves for this report; timings are the tester's, attributed to the thread.
Our verdict
The fortnight's story is one model: MiniMax H3 Image went from private to within 20% of the busiest open video model in five days, carried by a ComfyUI repack and a stock template. Under it, two quieter signals: base models are being downloaded again over their Turbo twins, and Civitai creators are moving LoRAs to Qwen 2.1 and H3. For a recurring face, the downloads say "try it"; the threads and the license say "do not switch yet". The next reading is due on 29 November.
Questions
- What is the most downloaded new image model on Hugging Face in November 2026?
- MiniMax H3 Image. On 15 November 2026 the Comfy-Org repack had 1,615,613 downloads since its creation on 8 November and the official MiniMaxAI repository had 517,948. Counts are file downloads, not users.
- Why does the Comfy-Org repack have more downloads than the official model?
- ComfyUI's stock templates point to the Comfy-Org files, so everyone who opens the template fetches them. Hugging Face counts downloads per repository, so the repack outranks the original even though it carries the same license.
- Is LTX-2.5 still the most downloaded open image-to-video model?
- Yes, in our 15 November 2026 reading: 2,777,667 downloads in the last 30 days, up 13.8% from 2,441,077 on 1 November.
- Can MiniMax H3 Image run on a 12 GB GPU?
- One tester reports it does with the int8 files and 96 GB of system RAM: 11 to 12 seconds per 4.1 MP image warm, with about 41 GB of RAM in use. That is one user's measurement, not ours.
- Can I use MiniMax H3 Image in the US or the EU?
- Not under the open-weight license. The MiniMax-H3-Image community license excludes the EU, the UK, the Republic of Korea and the USA; MiniMax's own Q&A says those regions need separate authorization.
Sources
- 01Comfy-Org/MiniMax-H3-Image · Comfy-Org, Hugging Face, 2026-11-15
- 02MiniMaxAI/MiniMax-H3-Image model card and license · MiniMax, Hugging Face, 2026-11-10
- 03PrunaAI/P-H3-Image-Comfy · Pruna AI, Hugging Face, 2026-11-15
- 04LTX-2.5 model card · Lightricks, Hugging Face, 2026-11-15
- 05Qwen-Image-2.1 model card · Qwen, Hugging Face, 2026-11-15
- 06Krea-2-Raw model card · Krea, Hugging Face, 2026-11-15
- 07Hugging Face trending text-to-image models · Hugging Face, 2026-11-15
- 0812 GB card numbers: int8 DiT + full 32B text encoder · Hugging Face community, 2026-11-11
- 09face consistency · Hugging Face community, 2026-11-11
- 10Opinion: not ready for an official release · Hugging Face community, 2026-11-11
- 11P-H3-Image vs official H3-Image on a 12 GB card · Hugging Face community, 2026-11-11
- 12Q&A About the License · MiniMax, Hugging Face, 2026-11-10
- 13Poor name · Hugging Face community, 2026-11-13
- 14MiniMax releases open-weight H3-Image model · Crypto Briefing
- 15Civitai most downloaded models · Civitai, 2026-11-15
- 16DaSiWa MiniMax H3 · Civitai, 2026-11-15
- 17Detailifier · Civitai, 2026-11-15
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.