AuraFlow¶

Development line: project:auraflow · thread auraflow-model-releases
Last event: 2024-08-15 · 3 dated since 2024-07-12 · Researched: 2026-09-04 · confidence: medium

What it is¶

A flow-based text-to-image model for local Diffusers and ComfyUI setups or hosted fal inference.

• Generates images from text through AuraFlowPipeline.
• Exposes a hosted fal-ai/aura-flow endpoint.
• v0.3 supports varied aspect ratios up to 1536 pixels per dimension.

Development line¶

• 2024-07-12 — AuraFlow's initial public project references were recorded. V0.1 launched as a fully open flow-based text-to-image model; fal documented native Diffusers and ComfyUI support and a 6.8B architecture.
• 2024-07-29 — AuraFlow v0.2 was recorded as a new model version. V0.2 was documented as a v0.1 successor trained with more compute. The official collection records its model update as 2024-07-27, so this may distinguish model publication from announcement.
• 2024-08-15 — AuraFlow v0.3 was recorded as a subsequent model version. V0.3 added more compute, aesthetic-data fine-tuning, and aspect-ratio support up to 1536 pixels per dimension.

What changed¶

AuraFlow — development line:

• 2024-07-12: v0.1 launched as a fully open flow-based text-to-image model; fal documented native Diffusers and ComfyUI support and a 6.8B architecture.
• 2024-07-29: v0.2 was documented as a v0.1 successor trained with more compute. The official collection records its model update as 2024-07-27, so this may distinguish model publication from announcement.
• 2024-08-15: v0.3 added more compute, aesthetic-data fine-tuning, and aspect-ratio support up to 1536 pixels per dimension.
• Found today: fal’s official collection lists only the unversioned model, v0.2, and v0.3; the live fal endpoint identifies AuraFlow as v0.3 and still labels it beta.

How to use this¶

As of 2024-08-15, treat AuraFlow as a versioned model line. Select the specific official AuraFlow, AuraFlow-v0.2, or AuraFlow-v0.3 reference needed for each workflow rather than assuming releases are interchangeable.

1. Pin fal/AuraFlow-v0.3 for a new local build; it is the documented current checkpoint in the official collection. — https://huggingface.co/fal/AuraFlow-v0.3
2. Install current diffusers, transformers, and accelerate, then load AuraFlowPipeline.from_pretrained("fal/AuraFlow-v0.3", torch_dtype=torch.float16, variant="fp16").to("cuda"). — https://huggingface.co/fal/AuraFlow-v0.3
3. Generate with explicit dimensions, 50 inference steps, CFG 3.5, and a fixed generator seed; save the resulting image. — https://huggingface.co/fal/AuraFlow-v0.3
4. For hosted inference, install @fal-ai/client, set FAL_KEY in the server environment, call fal.subscribe("fal-ai/aura-flow", { input: { prompt } }), and retain the returned request ID. — https://fal.ai/models/fal-ai/aura-flow/api

Best practices¶

• Set steps and CFG explicitly. Start from the v0.3 recipe of 50 steps and CFG 3.5 instead of relying on a library default. — https://huggingface.co/fal/AuraFlow-v0.3
• When local VRAM is constrained, use the documented 8-bit quantization and balanced device mapping, then compare outputs before adopting the lower-precision path. — https://huggingface.co/docs/diffusers/api/pipelines/aura_flow
• Keep FAL_KEY server-side; use queue status and webhooks rather than blocking for slow requests, and retain the returned seed and expanded prompt for reproducibility. — https://fal.ai/models/fal-ai/aura-flow/api
• Do not import the published v0.3 ComfyUI workflow unchanged: it currently names aura_flow_0.2.safetensors; replace the checkpoint deliberately and validate a test generation. — https://huggingface.co/fal/AuraFlow-v0.3/blob/main/comfy_workflow.json
• For SimpleTuner fine-tuning, treat its AuraFlow-specific Lycoris LoKr route as the starting point rather than casual full-rank tuning; its guide calls for substantial VRAM/DeepSpeed for full tuning and starts validation at 1024², 30–50 steps, and CFG 3.5–4.0. — https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/AURAFLOW.md

Superseded by this¶

• 2024-07-29: AuraFlow v0.2 supersedes the v0.1 checkpoint for the official mainline because it was trained with more compute; retain v0.1 only for legacy reproducibility.
• 2024-08-15: AuraFlow v0.3 supersedes v0.2 as the documented starting checkpoint, adding aesthetic fine-tuning and variable aspect ratios up to 1536 pixels per dimension.
• Current: unversioned fal/AuraFlow and v0.2 loading snippets are legacy for a new build; pin fal/AuraFlow-v0.3 instead. This is a usage recommendation, not a formal deprecation notice.

Still unknown¶

• fal’s collection lists three v0.x checkpoints, but that absence cannot prove no unpublished or differently named successor exists.
• The public model and API pages still call v0.3 beta; no current supported-GPU matrix or owned v0.3 run was found, so VRAM and latency are not generalized.
• The current Diffusers reference is internally inconsistent: its call signature shows CFG 3.5 while parameter prose says 5.0. Set CFG explicitly from the v0.3 card/API recipe.
• The v0.3 repository’s published comfy_workflow.json names a v0.2 checkpoint, so it is not verified as a v0.3 workflow without local replacement and a test run.
• The Simplified-Chinese lane found a SimpleTuner translation and one community ComfyUI workflow, but no independent, version-locked v0.3 operating measurement; it does not change the documented starting recipe.

Sources¶

Agent brief¶

• Subject: project:auraflow, thread auraflow-model-releases, 3 dated events 2024-07-12 → 2024-08-15.
• Practical note: As of 2024-08-15, practitioners should treat AuraFlow as a versioned model line and select the specific official AuraFlow, AuraFlow-v0.2, or AuraFlow-v0.3 reference needed for their workflow rather than assuming releases are interchangeable.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.