Best Models for ComfyUI in 2026

The model you load into ComfyUI determines most of what your output looks like — image quality, style range, resolution, and speed. In 2026 the landscape has shifted significantly: Flux.1 has raised the quality ceiling, but older models like SDXL and SD 1.5 still have genuine roles depending on your hardware and use case. This is a practical rundown of what’s worth using, and when.

Flux.1-dev

Best for: Highest-quality output for personal projects
VRAM requirement: 12GB+ (fp8), 16–24GB+ (full precision)
Licence: Non-commercial only

Flux.1-dev from Black Forest Labs is currently the best locally-runnable image generation model. It produces significantly better prompt adherence than SDXL — complex compositions, accurate text rendering (within limits), and sharp detail at native 1024px resolution. The step count matters: at 20–25 steps with CFG 1.0 and the euler sampler, results are consistently strong.

The trade-off is hardware: full-precision Flux wants 16–24GB VRAM. On 12GB cards, fp8 versions are the practical choice — quality is close, generation is a bit slower. If you have a high-end consumer GPU (RTX 3090, 4080, 4090) or a workstation card, this is the model to use.

Where to download: huggingface.co/black-forest-labs/FLUX.1-dev

Flux.1-schnell

Best for: Fast iteration, commercial projects
VRAM requirement: 8GB (fp8), 12GB+ (full precision)
Licence: Apache 2.0

Flux.1-schnell is a distilled version of flux1-dev that generates good results in 4–8 steps rather than 20+. The quality gap versus dev is real but smaller than you’d expect at typical output resolutions. For rapid prototyping, testing prompts, or any situation where you’re generating many images to find the right one, schnell’s speed advantage is significant.

The Apache 2.0 licence also makes it the right choice for any commercial or public-facing use where flux1-dev’s non-commercial restriction would be a problem.

Where to download: huggingface.co/black-forest-labs/FLUX.1-schnell

SDXL (Stable Diffusion XL)

Best for: General-purpose use, large fine-tuned model library
VRAM requirement: 8GB (base model, moderate settings)
Licence: OpenRAIL++ (permissive)

SDXL was released by Stability AI in 2023 and remains highly relevant. It generates at 1024×1024 natively, handles a wider style range than SD 1.5, and has an enormous library of fine-tuned checkpoints and LoRAs on Civitai.

The SDXL architecture has become the base for some of the most useful specialised models — Pony Diffusion (see below), Juggernaut XL for photorealism, and various illustration-focused fine-tunes. If you want to explore fine-tuned models beyond vanilla checkpoints, most of the interesting ones are SDXL-based.

SDXL also has a refiner model (sd_xl_refiner_1.0.safetensors) that can be used in a two-stage workflow for added detail — though in practice many users skip it and use an upscaler instead.

Where to download: huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 (base model). Fine-tunes: Civitai.

Pony Diffusion V6 XL

Best for: Stylised art, anime, illustration, character art
VRAM requirement: 8GB
Licence: Pony-specific terms (check Civitai listing)

Pony Diffusion is an SDXL-based fine-tune with a training dataset heavily weighted towards stylised art, illustrations, and character-focused content. It uses a tag-based prompting system (similar to Danbooru tags) rather than natural language sentences, which takes some adjustment if you’re used to other models.

For anything in the stylised, illustrative, or anime space, Pony is the go-to choice in 2026. Photorealism isn’t its strength — use Juggernaut XL or a Flux model for that — but for character illustration and stylised art, it’s the most capable SDXL fine-tune available.

Where to download: Civitai — search for “Pony Diffusion V6 XL”

Stable Diffusion 1.5

Best for: Low VRAM systems, fast iteration, the widest LoRA library
VRAM requirement: 4GB
Licence: OpenRAIL-M

SD 1.5 is the oldest of the mainstream models and shows it in output quality — 512×512 native resolution, softer detail, less prompt coherence than SDXL or Flux. But it still has a legitimate role:

• It runs on 4GB VRAM where nothing newer will
• It generates extremely quickly on any modern hardware
• It has the largest LoRA library of any model — thousands of fine-tuned styles and subjects on Civitai

If you’re testing a new custom node, debugging a workflow, or working with a LoRA that only exists for SD 1.5, it’s still useful. It’s also the right choice if your GPU has 4–6GB VRAM and you need to generate images at all.

Where to download: huggingface.co/runwayml/stable-diffusion-v1-5 or Civitai

Choosing Based on VRAM

• 4GB VRAM: SD 1.5 only. Small SDXL generations with --lowvram are possible but painful.
• 8GB VRAM: SDXL comfortably. Flux.1-schnell with fp8/GGUF quantisation. SD 1.5 very fast.
• 12GB VRAM: Flux.1-dev fp8, SDXL full precision, all SD 1.5/Pony variants.
• 16GB+ VRAM: Full-precision Flux.1-dev/schnell. No compromises on any current model.

Supporting Files

Models often need supporting files:

• VAE: Flux requires its own VAE (ae.safetensors). SDXL works best with the SDXL VAE. SD 1.5 has a built-in VAE but the vae-ft-mse-840000-ema-pruned is a popular improvement.
• LoRAs: Go in models/loras/. Work across architectures their fine-tune supports.
• Embeddings: Go in models/embeddings/. Mostly SD 1.5 and SDXL.

Where to Find Models

• Hugging Face — official model releases from Black Forest Labs, Stability AI, and research groups
• Civitai — community fine-tunes, LoRAs, and checkpoints. The largest library by volume.

For detailed Flux setup, see How to Install Flux in ComfyUI. For hardware recommendations based on these requirements, see Best GPU for ComfyUI in 2026.

Quick Recommendation by Use Case

For photorealistic images with the best current quality: Flux.1 Dev if your GPU has 12 GB+ VRAM, or Flux.1 Schnell for speed and commercial use. For stylised, illustrated, or anime content with broad community support: SDXL with a fine-tuned checkpoint from Civitai. For maximum compatibility with older workflows, LoRAs, and ControlNet setups: Stable Diffusion 1.5. For NSFW or stylised character art: Pony Diffusion XL. The single best all-round starting point for a new ComfyUI user with a modern GPU is Flux.1 Schnell — it requires no prompt engineering expertise to produce impressive results.

Fine-Tuned Checkpoints vs Base Models

The models listed above are all base models. Civitai hosts thousands of fine-tuned versions of each — trained on specific art styles, characters, or subject matter. A fine-tuned SDXL checkpoint trained on studio photography, for example, will produce more consistent professional-looking results for that use case than the base SDXL model. When you know the style or subject matter you want to focus on, searching Civitai for a fine-tuned version of whichever base model you are using is usually worth doing before spending time on prompt engineering.

For a full index of every ComfyUI guide on Serverman, see the ComfyUI complete guide and hub.