13 open models,
every file sized.

Text-to-image model by Tongyi-MAI (Alibaba). 6B parameters, S3-DiT architecture with 8-step turbo generation. Uses Qwen 3 4B as text encoder. Apache 2.0.

Compact image editing models by Black Forest Labs. 4B and 9B parameters, fast inference. Edit with natural language — backgrounds, clothing, photo restoration. The 4B is Apache 2.0.

Text-to-image model by Ideogram. 9.3B params, best-in-class typography and text rendering among open weights. Single-stream DiT paired with a Qwen3-VL-8B text encoder, loaded as conditional + unconditional models. Day-0 ComfyUI support. Non-commercial license.

Video + audio generation model by Lightricks. 21B parameters, Diffusion Transformer architecture. Text-to-video and image-to-video, synchronized audio, up to 4K. LTX-2 Community License.

Base (undistilled) 6B text-to-image model by Tongyi-MAI (Alibaba). Real CFG and negative prompts, more facial diversity — the model to fine-tune and train LoRAs on; pair with the turbo distill patch LoRA for 8-step inference. Shares text encoder and VAE with Z Image Turbo.

FLUX.1 Schnell is a 12B rectified-flow transformer for text-to-image, timestep-distilled to render high-quality images in just 1-4 steps. Released by Black Forest Labs under Apache 2.0 - the fully open, commercially usable member of the FLUX.1 family. Runs at guidance 1 with no negative prompt. The FP8 build is an all-in-one checkpoint (diffusion + text encoders + VAE bundled); BF16 and GGUF use separate components.

Qwen Image is a 20B-parameter MMDiT (Multimodal Diffusion Transformer) text-to-image model from Alibaba's Qwen team, released August 2025 under Apache 2.0. Its standout capability is state-of-the-art text rendering - multi-line layouts, paragraphs and fine print - in both alphabetic and logographic scripts (especially Chinese). It pairs a Qwen2.5-VL 7B text encoder with a custom VAE and generates natively at 1328x1328. A refreshed 2512 revision ships in the same Comfy-Org repo.

Image editing model by Alibaba's Qwen team, version 2511. 20B MMDiT with native multi-image editing (up to 3 reference images in one edit), improved facial identity preservation and instruction following. Shares text encoder and VAE with Qwen Image.

Image-to-video and text-to-video by Alibaba's Wan team. A14B is a mixture-of-experts pair: the high-noise expert authors composition, motion and camera; the low-noise expert resolves texture, faces and fine detail. The router is the timestep, so both experts are required. 81 frames at 16 fps native (5.06 s), 720x1280 or 480x832. Still the latest open-weights Wan — 2.5 and later are API-only.

Krea 2 (Turbo) is the 8-step distilled variant of Krea AI's Krea 2, a 13B-parameter dense Diffusion Transformer text-to-image model. It generates in roughly two seconds, pairing the Qwen Image VAE with a Qwen3-VL text encoder that uses multi-layer feature aggregation, and natively renders 1K-2K resolution (up to 2 MP). Distributed under the Krea 2 Community License — per Krea's announcement, commercial use is free for teams under 50 seats; see the license for binding terms.

Character animation model by Alibaba's Wan team. Reference image + driving video: the character performs the full motion and expression of the video, with viewpoint and background controlled by prompt. Native ComfyUI support, self-contained repo, 832x480 in 81-frame chainable segments.

Video + synchronized audio generation by Lightricks, successor to LTX-2.3. 22B DiT with Pixel Diffusion, native multishot (character and voice consistency across cuts), Gemma 4 12B text encoder and Auto Duration. Distilled checkpoint runs at 8 steps, CFG 1. Gated: accept the license on Hugging Face before downloading.

Omni-modal video generation by MiniMax (Hailuo). 33B dense DiT that generates video with native synchronized stereo audio (speech, SFX, music) in the same pass. Two checkpoints: FL2VA (t2v / i2v / first-last frame) and Ref2VA (up to 9 reference images + 3 videos + 3 audios lock identity, style and voice). 768p native, 24 fps, up to 15 s. Pruned variants (~20B) cut disk and RAM at equal output.