There have never been more open text-to-image models you can run at home. In 2026 alone we got Qwen-Image 2.1, Krea 2, FLUX 3 Image Klein, Z-Image, MOSS-Image and Fibo 2.0, and each one has its own licence, hardware needs and set of LoRAs. A model is only as useful as the styles, characters and speed-ups you can add to it, so this guide ranks the best open source text-to-image models of 2026 and gives you the best LoRAs for each one, with direct links. For every model you also get the exact ComfyUI files and folders.
- Best quality: Qwen-Image 2.1, the top open model on the Artificial Analysis arena (October 2026). It is for research and non-commercial use only.
- Best for 8 GB GPUs: FLUX 3 Image Klein (5.6B). It is free for commercial use under $5M revenue.
- Most LoRAs: Krea 2 (933 on Hugging Face) and Z-Image Turbo (838). Z-Image is Apache 2.0.
- Best fully permissive picks: Z-Image Turbo, Qwen-Image-2512, MOSS-Image and FLUX.2 klein 4B, all Apache 2.0.
- LoRA rule: a LoRA only works on the model family it was trained for. Train on the Base or Raw version and run it on the Turbo or distilled version.
Quick Comparison
| Model | Size | Licence | Commercial use | Practical VRAM | LoRAs on HF |
|---|---|---|---|---|---|
| Qwen-Image 2.1 | 7B generator | Qwen Research | No | 16 GB (INT8) | 7 |
| FLUX 3 Image Klein | 5.6B | FLUX Kommunity | Under $5M revenue | 8 GB (INT8) | 31 |
| Krea 2 | 12B | Krea 2 Community | Under 50 seats | 12 to 16 GB (fp8/nvfp4) | 933 |
| Z-Image Turbo | 6B | Apache 2.0 | Yes | 16 GB | 838 (+146 base) |
| Qwen-Image-2512 | ~20B | Apache 2.0 | Yes | 24 GB (fp8) | 200 |
| MOSS-Image | 5B | Apache 2.0 | Yes | 16 GB (INT8) | 96 |
| FLUX.2 klein 4B | 4B | Apache 2.0 | Yes | 8 to 13 GB | 112 |
| Fibo 2.0 | 8B | CC BY-NC 4.0 | No | 16 GB (INT8) | 3 |
| SDXL 1.0 | 3.5B | Open RAIL++-M | Yes | 8 GB | ~9,800 |
The LoRA counts are the number of adapter repos tagged with each base model on Hugging Face. Many more LoRAs live on Civitai, but Hugging Face counts are a fair way to compare ecosystems. Licences come from each model card. Always read the card yourself before you sell images.
1. Qwen-Image 2.1: Best Overall Quality
Alibaba’s Qwen-Image 2.1, released on September 20, 2026, is the highest-ranked open-weights model on the Artificial Analysis text-to-image arena, with an Elo of 1,035 in early October. It is much smaller than the original 20B Qwen-Image, with a 7B generator, and it does generation and editing in one checkpoint, with up to 10 reference images and transparent RGBA output. The catch is the Qwen Research licence, which allows research and non-commercial use only. Its LoRA ecosystem is still tiny, but the speed-ups that matter are already here.
ComfyUI files (Comfy-Org/Qwen-Image-2.1): ComfyUI/models/diffusion_models/qwen_image_2.1_int8_convrot.safetensors (or the Turbo version), ComfyUI/models/text_encoders/qwen3vl_8b_int8_convrot.safetensors, ComfyUI/models/vae/qwen_image_2.1_vae_bf16.safetensors. The full 16 GB setup is in my Qwen-Image 2.1 Turbo workflow.
| LoRA | What it does |
|---|---|
Turbo LoRA (official, in the Comfy repo: ComfyUI/models/loras/qwen_image_2.1_turbo_lora_avg_rank_178_bf16.safetensors) |
8 steps at CFG 1, applied to the base model |
| linoyts/Qwen-Image-2.1-DMD2-2steps | 2-step generation and editing at CFG 1 (strip the “transformer.” key prefix for ComfyUI) |
| Viggle/Qwen-Image-2.1-viggle-turbo | Community turbo model, also as a 4-step ComfyUI LoRA (t8star) |
| WarmBloodAban/Qwen-Image-2.1-LoRAs | A collection of community style LoRAs |
| e-n-v-y/Qwen-Image-2.1-Fix | Quality fix LoRA for common artefacts |
2. FLUX 3 Image Klein: Best for 8 GB GPUs
Black Forest Labs released FLUX 3 Image Klein on October 15, 2026: a 5.6B image transformer with a Qwen3-VL-4B text encoder, native 4K output and editing built in. BFL says the INT8 build runs on GeForce RTX cards with as little as 8 GB of VRAM, which makes it the strongest new model for mid-range PCs. Under the FLUX Kommunity Licence, commercial use is free for individuals and companies under $5M in annual revenue. Use Klein Base (50 steps) for LoRA training, and the distilled Klein (16 steps) for everyday generation.
ComfyUI files (black-forest-labs/flux-3-image-klein-comfy, gated): ComfyUI/models/diffusion_models/flux3-image-klein-int8-convrot.safetensors, ComfyUI/models/text_encoders/qwen3vl_4b_int8_convrot.safetensors, ComfyUI/models/vae/flux2-vae.safetensors. Details in my FLUX 3 Image Klein guide.
| LoRA | What it does |
|---|---|
| Muinez/flux3-klein-distil-4step | 4-step distillation for the Base model |
| NoiseToNova/Flux-3-Klein-turbo-pdd-4_8-steps | Turbo LoRA for 4 or 8 steps |
| ausboss/FLUX-3-Image-Klein-Outpaint-LoRA | Outpainting (extend an image beyond its borders) |
| ausboss/FLUX-3-Image-Klein-Consistency-LoRA | Keeps subjects consistent across edits |
| ausboss/FLUX-3-Image-Klein-Restore-LoRA | Photo restoration |
| akhaliq/flux-3-image-klein-angles-lora | New camera angles of the same subject |
3. Krea 2: Biggest New LoRA Ecosystem
Krea 2 is a 12B model trained from scratch by Krea AI, with open weights released in June 2026. It comes as Raw, an undistilled base for LoRA training, and Turbo, which makes 2K images in 8 steps with CFG off. Its look is the most varied and “artistic” of the group, and it reached the top 10 of the Artificial Analysis leaderboard. With 933 adapters on Hugging Face in four months, it is the fastest-growing LoRA scene of 2026. The Krea 2 Community Licence allows free commercial use for teams under 50 seats, with a revenue cap and required content filtering.
ComfyUI files (Comfy-Org/Krea-2): ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors (or krea2_turbo_nvfp4 on RTX 50 series), ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors, ComfyUI/models/vae/qwen_image_vae.safetensors. Comfy’s advice: train LoRAs on Raw and run them on Turbo.
| LoRA | What it does |
|---|---|
10 official style LoRAs in Comfy-Org/Krea-2 (ComfyUI/models/loras/krea2_softwatercolor.safetensors, krea2_retroanime, krea2_vintagetarot, krea2_darkbrush and more) |
Styles with trigger words, for example “art deco watercolor style”, strength 1.0 |
| conradlocke/krea2-identity-edit | Identity-preserving edits (the most liked Krea 2 adapter) |
| gokaygokay/Krea-2-Realism-LoRA | More photographic realism |
| AliveAi/Krea-2-Edit-Outfit-Transfer | Move an outfit from one image to a person |
| Patil/Krea-2-depth-controlnet | Depth map control |
| TheDivergentAI/krea2-turbo-distill-lora | Turbo speed as a LoRA on Raw |
| e-n-v-y/Krea-2-Pixel-Art | Pixel art style |
4. Z-Image Turbo and Z-Image: Best Apache 2.0 All-Rounder
Z-Image from Alibaba’s Tongyi-MAI lab is a 6B model with a fully permissive Apache 2.0 licence. Z-Image Turbo needs only 8 function evaluations per image and “fits comfortably within 16G VRAM consumer devices”, according to its card. It is known for realistic photos and clean text in images. The Turbo card says it is not meant for fine-tuning, so train on the Z-Image base. Even so, Turbo has 838 LoRAs, the second biggest set in this list.
ComfyUI files (Comfy-Org/Z-Image-Turbo): ComfyUI/models/diffusion_models/z_image_turbo_int8_convrot.safetensors (bf16 also available), with the Qwen3-4B encoder in ComfyUI/models/text_encoders/ and the VAE in ComfyUI/models/vae/, as listed in the repo’s split_files folder.
| LoRA | What it does |
|---|---|
| Shakker-Labs/AWPortrait-Z | High-end portrait look (the most liked Z-Image LoRA) |
| ostris/zimage_turbo_training_adapter | Lets you train LoRAs directly for Turbo in AI Toolkit |
| renderartist/Technically-Color-Z-Image-Turbo | Vintage Technicolor film look |
| tarn59/pixel_art_style_lora_z_image_turbo | Pixel art style |
| wcde/Z-Image-Turbo-DeJPEG-Lora | Removes JPEG-like compression artefacts |
| suayptalha/Z-Image-Turbo-Realism-LoRA | Extra photorealism |
| nphSi/Z-Image-Lora (base) | The most downloaded LoRA set for Z-Image base |
5. Qwen-Image-2512: Best Big Model You Can Sell With
If you like Qwen’s prompt following and text rendering but need a commercial licence, use Qwen-Image-2512, the December 2025 update of the 20B Qwen-Image under Apache 2.0. It scored 999 on the arena, just behind FLUX.2 dev, and is known for realistic people and fine detail. The trade-off is size: Comfy’s fp8 test used 86% of a 24 GB card, so it is best for 24 GB GPUs or with heavy offloading on 16 GB.
ComfyUI files: from Comfy-Org/Qwen-Image, with the transformer in ComfyUI/models/diffusion_models/, the Qwen2.5-VL-7B encoder in ComfyUI/models/text_encoders/ and the Qwen-Image VAE in ComfyUI/models/vae/.
| LoRA | What it does |
|---|---|
| lightx2v/Qwen-Image-2512-Lightning | Few-step Lightning speed-up (132k downloads) |
| Wuli-art/Qwen-Image-2512-Turbo-LoRA | Turbo LoRA, plus a 2-step version |
| Danrisi/Samsung_Qwen2512 | Phone-camera photo realism |
| prithivMLmods/Qwen-Image-2512-Pixel-Art-LoRA | Pixel art style |
6. MOSS-Image and Kromoss: Best for Layout Control
MOSS-Image from OpenMOSS is a 5B transformer guided by a Qwen3.5-9B language model, under Apache 2.0. Its stand-out feature is JSON layout prompts with bounding boxes, and it is strong with tag-style prompts. In the last two weeks the community added speed LoRAs, an edit LoRA and a full fine-tune called Kromoss. The INT8 build fits a 16 GB card.
ComfyUI files (Comfy-Org/MOSS-Image): ComfyUI/models/diffusion_models/moss_image_sft_int8_convrot.safetensors, ComfyUI/models/text_encoders/qwen3.5_9b_int8_convrot.safetensors, ComfyUI/models/vae/flux2-vae.safetensors. See my MOSS-Image guide.
| LoRA | What it does |
|---|---|
| darask0/MOSS-Image_4-step_adapter_lora | 4-step generation |
| lodestones/MOSS-Image-SFT-8-steps-dmd | 8-step DMD speed-up |
| chinoll/Moss-Image-Edit | Instruction editing with up to 4 references (my guide) |
| darask0/MOSS-Image-color-depth-control | Colour-region and depth map control, supported in ComfyUI core |
| paralaif/MOSS-Image-Alpha-LoRA | Transparent-background output |
| MATLOWAI/MOSS-Image-SFT-DeBand | Removes banding and stripes in skies |
| 7err4/MOSS-Image-CharRef | Character reference for consistent characters |
7. FLUX.2 klein 4B: Best Small Commercial Model
FLUX.2 [klein] 4B, from January 2026, is the only Black Forest Labs model with a plain Apache 2.0 licence. It handles generation and multi-reference editing, and Comfy measured about 8.4 GB of VRAM for the distilled version. FLUX 3 Klein beats it on quality, but if you need zero licence worries on a small GPU, this is the one. Train on the base model.
ComfyUI files (Comfy-Org/Flux-2-Klein-4B): ComfyUI/models/diffusion_models/flux-2-klein-4b.safetensors (or flux-2-klein-base-4b.safetensors), ComfyUI/models/text_encoders/qwen_3_4b.safetensors, ComfyUI/models/vae/flux2-vae.safetensors.
| LoRA | What it does |
|---|---|
| svntax-dev/pixel_spritesheet_4walk_small_lora_v1 | Pixel art walk-cycle sprite sheets for games |
| nomadoor/flux-2-klein-4B-360-erp-outpaint-lora | 360-degree panorama outpainting |
| trmz/flux2-klein-alpha | Transparent-background images |
| ostris/sketch_to_image_klein_4b | Turns sketches into finished images |
| thedeoxen/refcontrol depth and xocialize/refcontrol pose | Depth and pose control from a reference |
8. Bria Fibo 2.0: Best Precise Control
Bria’s Fibo 2.0 is an 8B model trained only on licensed data, and it is driven by structured JSON prompts that set the camera, lens, lighting and depth of field. Its Edit Turbo checkpoint is one of the best local image editors right now. The weights are CC BY-NC 4.0, so commercial use needs a deal with Bria. LoRA support is young: Bria says the Dev checkpoint is the base for LoRAs, and only a handful exist so far.
ComfyUI files (Comfy-Org/Fibo-2.0): ComfyUI/models/diffusion_models/fibo2_dev_int8_convrot.safetensors (or fibo2_edit_turbo_int8_convrot.safetensors), ComfyUI/models/text_encoders/qwen3vl_4b_int8_convrot.safetensors, ComfyUI/models/vae/flux2-vae.safetensors. See my Fibo 2.0 guide.
| LoRA | What it does |
|---|---|
| akhaliq/Fibo-2.0-8step-smd | 8-step speed-up for Dev |
| axiomofmind/Fibo-2.0-anime-lora | Anime style |
| axiomofmind/Fibo-2.0-psychedelic-lora | Psychedelic art style |
9. SDXL and Its Fine-Tunes: Biggest LoRA Library
SDXL is from 2023 and its raw quality is well behind everything above, but nothing comes close to its ecosystem: about 9,800 adapters on Hugging Face and many more on Civitai, plus ControlNets for everything. It runs on 8 GB cards, and fine-tunes like Illustrious-XL and NoobAI-XL (anime) or RealVisXL (photo) are still the standard for anime and character work. Check each fine-tune’s licence; NoobAI-XL, for example, forbids commercial use.
ComfyUI files: one checkpoint file, ComfyUI/models/checkpoints/sd_xl_base_1.0.safetensors (or your fine-tune), loaded with Load Checkpoint.
| LoRA | What it does |
|---|---|
| latent-consistency/lcm-lora-sdxl | 4 to 8 step LCM speed-up |
| h1t/TCD-SDXL-LoRA | Few-step TCD speed-up with more detail than LCM |
| nerijs/pixel-art-xl | Pixel art |
| alvdansen/littletinies | Cute hand-drawn illustration style |
| ostris/ikea-instructions-lora-sdxl | IKEA instruction-manual drawings |
| goofyai/3d_render_style_xl | 3D render style |
| artificialguybr/StickersRedmond | Sticker designs |
How to Use These LoRAs in ComfyUI
- Download the
.safetensorsfile from the LoRA’s Hugging Face page (Files tab) and put it inComfyUI/models/loras/. A subfolder per model, for exampleComfyUI/models/loras/krea2/, keeps things tidy. - Add a LoraLoaderModelOnly node between the model loader and the sampler (SDXL uses the normal Load LoRA node, which also patches the text encoder).
- Start at strength 1.0 and lower it if the style is too strong. Add the trigger word from the model card to your prompt.
- Match the LoRA to the exact model family. A Krea 2 LoRA will not work on Z-Image, and a Qwen-Image-2512 LoRA will not work on Qwen-Image 2.1.
- For speed LoRAs, set the steps and CFG from the LoRA card (for example 8 steps at CFG 1 for the Qwen 2.1 Turbo LoRA) and do not stack two speed LoRAs.
Some packs, such as diffusers-format LoRAs, need their key names converted for ComfyUI. The card usually says so.
Which Model Should You Pick?
- Best images, personal use: Qwen-Image 2.1.
- 8 GB GPU: FLUX 3 Image Klein, or FLUX.2 klein 4B if you need Apache 2.0.
- 16 GB GPU and client work: Z-Image Turbo or MOSS-Image.
- Most styles and characters: Krea 2, or SDXL for anime.
- Exact camera and lighting control: Fibo 2.0.
- 24 GB GPU, commercial, top quality: Qwen-Image-2512.
Not on the list: FLUX.2 [dev] (32B) and HunyuanImage 3.0 (80B) are too big for home PCs, Ideogram 4 needs 24 GB and is non-commercial, and HiDream-O1 and ERNIE-Image are good but have almost no LoRAs yet.
FAQ
What is the best open source text-to-image model in 2026?
For pure quality, Qwen-Image 2.1, the top open model on the Artificial Analysis arena in October 2026. For a fully open licence, Z-Image Turbo or Qwen-Image-2512, both Apache 2.0.
Which open image models can I use commercially?
Z-Image, Qwen-Image-2512, MOSS-Image, FLUX.2 klein 4B and SDXL allow commercial use. FLUX 3 Klein (under $5M revenue) and Krea 2 (under 50 seats) allow it with limits. Qwen-Image 2.1 and Fibo 2.0 are non-commercial.
Can I use a LoRA from one model on another?
No. LoRAs only work on the model family they were trained for. Within a family, LoRAs trained on Base or Raw usually work on the Turbo or distilled version.
Which model is best for an 8 GB GPU?
FLUX 3 Image Klein in INT8, FLUX.2 klein 4B, or SDXL.
Where do LoRA files go in ComfyUI?
In ComfyUI/models/loras/. Diffusion models go in ComfyUI/models/diffusion_models/, text encoders in ComfyUI/models/text_encoders/ and VAEs in ComfyUI/models/vae/.
Final Thoughts
The best open source text-to-image model depends on three things: your GPU, your licence needs and the LoRAs you want. Qwen-Image 2.1 has the best raw quality, FLUX 3 Klein is the best new model for small cards, Krea 2 and Z-Image have the busiest LoRA scenes, and MOSS-Image and Fibo 2.0 give you control the others cannot. Pick one, add a speed LoRA, and build from there.
More guides:
- Best Local AI Video Models in 2026
- How Much VRAM Do You Need for ComfyUI?
- Train a MOSS-Image LoRA With KohakuMoss
Sources
- Pinggy: Best open-source AI image models ranked (Arena data, October 7, 2026)
- iimagined: Best ComfyUI models 2026 (licences and VRAM, October 11, 2026)
- Hugging Face: FLUX 3 Image Klein, Comfy-Org/Krea-2, Comfy-Org/Qwen-Image-2.1, Comfy-Org/Flux-2-Klein-4B, Comfy-Org/Fibo-2.0
- Hugging Face adapter listings for each base model (counts checked October 29, 2026)
- ComfyUI Wiki: Qwen-Image 2.1 DMD2 2-step LoRA
- Krea 2 overview
