ModelMergeAuraflow

ModelMergeAuraflow is the block-merge node for AuraFlow, fal's 6.8B rectified-flow transformer - the Apache-2.0 model that answered the SD3 license collapse three weeks before Flux answered everything else. The KB's verdict is that AuraFlow is the clearest case of a model getting the license right and everything else too late: v0.2 genuinely beat Flux Dev on prompt adherence, then v0.3 traded it away, Flux landed, and AuraFlow never got another version.

So why does the node exist? Because a handful of people still run it - and one very specific person made it matter: AstraliteHeart chose AuraFlow as the base for Pony V7, for exactly the reason the model was built, that you can use it commercially without asking anyone. If you're working with the Pony V7 line or the tiny AuraFlow checkpoint ecosystem (seven resources on CivitAI carry it as their base), this is your merge tool. For everyone else, it's a museum piece with working controls - and that's okay.

How it works

Standard family mechanism - ModelMergeBlocks subclass, per-tensor interpolation result = ratio · model1 + (1 − ratio) · model2. AuraFlow's architecture is unusual, a mix of a DiT-ish core with a distinctive two-stage structure, which shows up in the key map:

• init_x_linear., final_linear. - input and output projections.
• positional_encoding, register_tokens - the positional encoding table and the learnable register tokens (yes, AuraFlow uses register tokens, which is why they get a slider).
• cond_seq_linear. - conditioning sequence projection.
• t_embedder. - timestep embedding.
• double_layers.0–3 - 4 double (joint image+text) attention blocks.
• single_layers.0–31 - 32 single attention blocks.
• modF. - the modulation final block.

All defaults at 1.0 (all model1). The structure is small enough that you could reasonably move individual ranges; double layers = early fusion, single layers = the bulk.

The inputs that matter

Two MODEL inputs plus the block sliders; there are no surprises beyond the architecture names. If you're blending two Pony V7-line checkpoints, the practical pattern is the usual structure-vs-detail split - early single layers and double layers for structure, late single layers for style. The register_tokens slider is the distinctive one; nobody outside AuraFlow has that knob, and it's a genuine "if you know, you know" control.

One MODEL out, into a KSampler or Save Checkpoint.

Where people get burned

AuraFlow's ecosystem is tiny and frozen - the last thread titled after it in the community corpus scored +1 with zero replies. Don't expect a body of merge recipes to lean on. Also be aware of the base-model reality: Pony V7 sat on a base with no trainers, no LoRAs, and a four-channel SDXL VAE, and AuraFlow's own history is a cautionary tale about version churn (v0.3 traded away the prompt adherence v0.2 was praised for). If you merge two AuraFlow checkpoints from different versions, you may be averaging contradictory behaviors.

How you get it

Core ComfyUI, model/merging/model specific, added November 2024 when AuraFlow support shipped. No installation.