Image models 28
We track 28 open image models from 8 labs, 13 of them released in the 12 months to 7 Nov 2026. The newest is MiniMax-H3-Image (3 Nov 2026) and the most downloaded is stable-diffusion-xl-base-1.0, with 4,747,305 downloads in 30 days. 22 fit a 16 GB GPU at 4-bit, and none has a hosted API in our data.
ComfyUI can load a complete image-generation checkpoint or assemble a pipeline from separate diffusion models, text encoders and VAEs. The distinction matters when choosing a download: a component checkpoint may need other weights and a matching workflow before it can produce an image.
What the category covers
Text-to-image generation is the headline use, but the same model families cover image editing, where an existing image plus an instruction produces a modified one, and inpainting, where part of an image is replaced. ComfyUI’s catalogue separates these, listing image generation workflows for model families including Stable Diffusion 1.5, SDXL, SD3.5 and Flux, and image editing workflows for editing-specific variants such as Flux Kontext and Qwen Image Edit.
Component models complicate the counting. A decoder or VAE released on its own is not a generator you can prompt, even though it appears in the same category.
How they are measured
Image generation has public evaluation tools, including GenEval, which tests object-focused properties of text-to-image output. Such a benchmark answers a defined question; it does not cover every aesthetic preference or editing requirement. Compare prompts and evaluation settings before comparing scores.
Run representative prompts across candidates and inspect the outputs as well as any benchmark scores. Test the features your job requires, such as object placement, text rendering or editing consistency. Video has its own evaluation suites, described on our video page.
What serves them
Diffusers and ComfyUI provide different interfaces for working with generation pipelines.
Diffusers is the library route. Hugging Face describes it as the go-to library for pretrained diffusion models for generating images, audio and even 3D structures of molecules, offering both simple inference and the pieces to train your own. It is Apache 2.0 licensed. If you are writing Python around generation, this is the layer you call.
ComfyUI is the graph route. Its README calls it a modular node graph interface for visual professionals who want control over every model, parameter and output, and claims to be the most optimised inference engine for diffusion models with a focus on local consumer hardware. That claim is the project’s own. Its workflow catalogue is the fastest way to see which model families are currently supported.
Core vLLM and llama.cpp are token generators: vLLM’s documentation says the generative models it supports share the ability to generate text, and the llama.cpp README describes the project as LLM inference in C/C++. For other outputs, vLLM’s documentation says vLLM-Omni “supports the generation of multimodal content, including images, videos, and audio”. Check that a runtime explicitly supports the model architecture and its required components before relying on it.
The hardware reality
Memory pressure comes from a different direction than in text. The weights are usually smaller than a large language model, but generation holds activations for the full latent at the target resolution, so requirements scale with image size and batch count rather than with context length. Pushing resolution up is the fastest way to exhaust a card.
Because the pipeline is assembled from parts, components can be placed separately: text encoders and VAEs can sit on CPU or be swapped in and out while the diffusion model stays resident. Quantised and reduced-precision releases of individual components reduce the memory each one needs.
Step count is the other lever. Distilled variants can reduce the number of denoising steps. That changes generation time and may affect output quality, but does not by itself prove a smaller weight-memory requirement.
Licence patterns
The runtime licence does not determine the checkpoint licence. Diffusers uses Apache 2.0, explained on our Apache 2.0 page, but a pipeline can load model components with different terms. Check each checkpoint and adapter you intend to use.
Do not infer a model’s permissions from another release by the same lab. A licence may set conditions on commercial use, redistribution or derivatives. Our MIT and modified MIT pages explain why a familiar label is not enough when additional terms apply.
This summarises licence text. It is not legal advice.
FAQ
Should I use diffusers or ComfyUI?
Diffusers if you are writing code around generation and want a stable Python API. ComfyUI if you want to assemble and tune pipelines visually, and it targets local consumer hardware explicitly.
Can I run image models on an AMD or Apple GPU?
Both projects support more than CUDA, and the practical answer depends on the specific pipeline. Our compute hub covers which stacks work on which hardware.
Why are there no benchmark tables on this page?
This article explains the tooling. Benchmarks such as GenEval test defined capabilities, while selecting a model also requires checking outputs on your own prompts.
Are commercial uses allowed?
Read the checkpoint’s licence. Open weights do not by themselves establish permission for commercial use, and the runtime’s licence cannot grant rights in separately supplied model weights.
Pick a view
All modelsNewest
| Model | Released | Type | Size | Memory at 4-bit | 16 GB GPU | Downloads, 30 days | Cheapest API, blended / 1M |
|---|---|---|---|---|---|---|---|
| 3 Nov 2026 | Image | Unknown | Unknown | Unknown | 437,039 | No tracked API | |
| 14 Oct 2026 | Image | Unknown | Unknown | Unknown | 1,330,289 | No tracked API | |
| 13 Oct 2026 | Image | 5.5B | 4 GB | Fits | 204,224 | No tracked API | |
| 13 Oct 2026 | Image | 5.5B | 3.9 GB | Fits | 135,059 | No tracked API | |
| 7 Apr 2026 | Image | 8B | 5.7 GB | Fits | 1,839 | No tracked API | |
| 6 Apr 2026 | Image | 62M | 0 GB | Fits | 161,275 | No tracked API | |
| 2 Apr 2026 | Image | 8B | 5.7 GB | Fits | 2,924 | No tracked API | |
| 14 Jan 2026 | Image | Unknown | Unknown | Unknown | 16,014 | No tracked API | |
| 14 Jan 2026 | Image | 3.9B | 2.8 GB | Fits | 438,415 | No tracked API | |
| 14 Jan 2026 | Image | 3.9B | 2.8 GB | Fits | 738,959 | No tracked API | |
| 14 Jan 2026 | Image | 9.1B | 6.5 GB | Fits | 65,956 | No tracked API | |
| 14 Jan 2026 | Image | 9.1B | 6.5 GB | Fits | 130,960 | No tracked API |
Most downloaded
| Model | Released | Type | Size | Memory at 4-bit | 16 GB GPU | Downloads, 30 days | Cheapest API, blended / 1M |
|---|---|---|---|---|---|---|---|
| 25 Jul 2023 | Image | 2.6B | 1.8 GB | Fits | 4,747,305 | No tracked API | |
| 14 Oct 2026 | Image | Unknown | Unknown | Unknown | 1,330,289 | No tracked API | |
| 14 Jan 2026 | Image | 3.9B | 2.8 GB | Fits | 738,959 | No tracked API | |
| 27 Nov 2023 | Image | 2.6B | 1.8 GB | Fits | 571,101 | No tracked API | |
| 31 Jul 2024 | Image | 11.9B | 8.5 GB | Fits | 482,190 | No tracked API | |
| 14 Jan 2026 | Image | 3.9B | 2.8 GB | Fits | 438,415 | No tracked API | |
| 3 Nov 2026 | Image | Unknown | Unknown | Unknown | 437,039 | No tracked API | |
| 27 Nov 2023 | Image | 866M | 0.6 GB | Fits | 413,273 | No tracked API | |
| 22 Nov 2025 | Image | 32.2B | 23 GB | Too big | 290,145 | No tracked API | |
| 31 Jul 2024 | Image | 11.9B | 8.5 GB | Fits | 274,082 | No tracked API | |
| 28 May 2025 | Image | 11.9B | 8.5 GB | Fits | 226,441 | No tracked API | |
| 13 Oct 2026 | Image | 5.5B | 4 GB | Fits | 204,224 | No tracked API |
Smallest that fit 16 GB
| Model | Released | Type | Size | Memory at 4-bit | 16 GB GPU | Downloads, 30 days | Cheapest API, blended / 1M |
|---|---|---|---|---|---|---|---|
| 6 Apr 2026 | Image | 62M | 0 GB | Fits | 161,275 | No tracked API | |
| 27 Nov 2023 | Image | 866M | 0.6 GB | Fits | 413,273 | No tracked API | |
| 11 Jul 2024 | Image | 1.3B | 0.9 GB | Fits | 161 | No tracked API | |
| 12 Jun 2024 | Image | 2.1B | 1.5 GB | Fits | 57,934 | No tracked API | |
| 26 Jul 2023 | Image | 2.3B | 1.6 GB | Fits | 110,642 | No tracked API | |
| 29 Oct 2024 | Image | 2.5B | 1.8 GB | Fits | 80,031 | No tracked API | |
| 25 Jul 2023 | Image | 2.6B | 1.8 GB | Fits | 4,747,305 | No tracked API | |
| 27 Nov 2023 | Image | 2.6B | 1.8 GB | Fits | 571,101 | No tracked API | |
| 14 Jan 2026 | Image | 3.9B | 2.8 GB | Fits | 738,959 | No tracked API | |
| 14 Jan 2026 | Image | 3.9B | 2.8 GB | Fits | 438,415 | No tracked API | |
| 13 Oct 2026 | Image | 5.5B | 3.9 GB | Fits | 135,059 | No tracked API | |
| 13 Oct 2026 | Image | 5.5B | 4 GB | Fits | 204,224 | No tracked API |
Releases per month
13 in 12 months| Month | Total |
|---|---|
| Dec 2025 | 1 |
| Jan 2026 | 5 |
| Feb 2026 | 0 |
| Mar 2026 | 0 |
| Apr 2026 | 3 |
| May 2026 | 0 |
| Jun 2026 | 0 |
| Jul 2026 | 0 |
| Aug 2026 | 0 |
| Sep 2026 | 0 |
| Oct 2026 | 3 |
| Nov 2026 | 1 |
Cheapest hosted APIs
All API pricesWe have not recorded API pricing for these image models yet.