Flux2TTRTrainingParameters

Flux2TTRTrainingParameters doesn't train anything itself - it's the shared config bag that both training phases read. You dial in the loss weights, optimizer settings, and Comet logging once, and it emits a TTR_TRAINING_CONFIG wire that feeds Flux2TTRControllerTrainer (required there) and Flux2TTRTrainer (optional there). One set of hyperparameters across Phase 1 and Phase 2 is the whole point: the distillation and the controller are trained against the same numbers, and you don't have to keep two workflows in sync by hand.

The inputs that matter

For a first run, the inputs that actually matter are a small handful:

• rmse_weight and cosine_weight - both default 1. These are the workhorse quality terms in controller training.
• Everything else starts at 0: lpips_weight, dreamsim_weight, hps_weight, biqa_quality_weight, biqa_aesthetic_weight, gemini_similarity_weight, gemini_quality_weight. Raising any of them pulls in that scorer - a heavy dependency (pyiqa, hpsv2, dreamsim) or, for the Gemini terms, the Google API.
• target_ttr_ratio (0.7), lambda_eff (1), lambda_entropy (0.1) - the reward-shaping knobs for the controller. target_ttr_ratio is the fraction of eligible layers you want routed to the fast TTR path.
• learning_rate (1e-4), grad_clip_norm (1), huber_beta (0.05) - optimizer and Phase-1 loss settings.
• comet_enabled - defaults to false, so you're not shipping runs to a cloud dashboard unless you flip it.

Gemini, only if you ask for it

The Gemini angle deserves a callout, because it's easy to misread. The node has gemini_model (default gemini-3-flash-preview), gemini_api_key, and gemini_api_key_env (default GEMINI_API_KEY), but with both gemini weights at zero the trainer never calls the API. You only need a key when you raise gemini_similarity_weight or gemini_quality_weight - and those calls cost real money, image-by-image. The key field is a plaintext string sitting in your workflow JSON, so think twice before sharing a workflow with a key baked in. Better: leave it empty and set the GEMINI_API_KEY env var.

Output is a single training_config (TTR_TRAINING_CONFIG), which is what you wire into the trainer nodes.

It's a config node, so there's nothing to get wrong at runtime - the failure modes live in whatever trainer you feed it to. The one genuine trap is the optional quality scorers. The pack's installer (via uv pip install -e custom_nodes/ComfyUI-Taylor-Attention, or bash install_requirements.sh /venv/main) handles pyiqa, hpsv2, and dreamsim, but image-reward is explicitly not installed by default because it pins an old timm that breaks pyiqa. If a tutorial tells you to add it, don't.

Installing and gotchas

Install the pack through ComfyUI Manager (search "Taylor-Attention") or clone it into custom_nodes, then restart and run the dependency install above. Note the README's one-liner still names the old folder ComfyUI-Approximate-Attention - use the folder you actually cloned. No model files are downloaded; the checkpoints end up under ComfyUI/models/approximate_attention/. This is a research pack with essentially zero community footprint and everything flagged experimental - treat the defaults as the author's best guess and tune from there.