AceStep 1.5 SFT Turbo Tag Adapter

Here's a situation you'll recognize if you've spent time around distilled image models: you grab a caption written for a Turbo model, paste it into a full-quality model, and get something that feels off - same semantic intent, wrong execution. AceStepSFTTurboTagAdapter exists because the same thing happens with music. It takes Turbo-style tags and rewrites them into shorter, SFT-friendly tags, then also hands you suggested cfg and steps values tuned for the SFT model. It's beta, it's a niche within a niche, and if you bounce between the AceStep Turbo and SFT checkpoints it's genuinely useful.

Why the mismatch exists

Turbo and SFT aren't just two versions of the same model - they're on opposite ends of the distillation spectrum. Turbo is a distilled model: it's trained to produce a full song at a handful of steps with very low (or zero) guidance, and its prompt dialect is tuned for that. SFT is the full supervised-fine-tuned model: it wants a conventional step count, CFG around 7, and reads tags differently. This is the same dynamic the image world runs into with SDXL Turbo vs a full checkpoint - the CFG and step expectations are baked into the model, not optional preferences.

The practical consequence: a Turbo caption that reads "Aggressive Brazilian Funk Mandelao, Ritualistic Phonk atmosphere, Heavy distorted 808 bass" is dense and stylistically loaded, and the SFT model chokes on or misreads that phrasing. The adapter's job is to translate the dialect, not the meaning.

How it works

It's a rule-based mapping, not an LLM call - instant, deterministic, no API key. Feed it turbo_tags (comma-separated works best) and pick an adaptation_strength:

• conservative - light rewording, keep most of the original phrasing.
• balanced (default) - the sensible middle.
• aggressive - heavy rewrite toward SFT-friendly short tags.

Two toggles refine the output: keep_unknown_tags (default true) preserves anything the mapper didn't recognize, after simplification; add_sft_bias_tags (default true) appends a few anchor tags for genre, groove, and mood that the SFT model responds to. It returns:

• sft_tags - the adapted comma-separated tags, ready for the TextEncode caption box.
• notes - a string explaining what changed.
• suggested_cfg (FLOAT) and suggested_steps (INT) - starting points for the Generate node's cfg and steps, based on the adaptation.

The right way to use it

Wire the outputs into your graph as primitives: sft_tags → TextEncode caption, and feed suggested_cfg/suggested_steps into Generate. Treat the suggestions as starting points, not gospel - the README's quality baseline for SFT is cfg 7.0 at 50 steps with euler/normal and APG guidance, and the adapter's numbers will usually land in that neighborhood.

Honest caveats: it's labeled BETA, it's mapping text to text so it can only be as good as its dictionary, and its real value shows up only when you're actively porting captions between the Turbo and SFT models. If you only ever use one checkpoint, you don't need it. If you're farming ideas on Turbo (fast, low steps) and then re-rendering the winners on SFT (slow, high quality) - which is the classic distilled-then-full workflow - this saves you from hand-editing every caption in the loop.

Install the pack once via ComfyUI Manager (search "AceStep") or git clone https://github.com/jeankassio/ComfyUI-AceStep_SFT.git into custom_nodes/. It has no model files and no heavy dependencies - pure string logic.