ZImage Turbo Single Prompt Conditioning Advanced

This is the version of ZImageTurboSinglePromptConditioning where the author stopped hiding the controls. Same one-prompt input, same three outputs, but instead of two fixed enhancement profiles you get every parameter of the basic and advanced enhancers as a widget. If the capitan_daily / capitan_literal presets were a locked door, this node is the keys.

What the two stacks do. The pipeline runs in two stages, and you can flip either on or off:

• Basic enhancer (use_basic_enhancer): the "glue" stage. Normalizes the embedding, pushes it through a small MLP with basic_mlp_hidden_mult hidden width, and blends it back at basic_strength. basic_normalize stabilizes the embedding before the MLP; basic_self_attention adds a self-attention pass that helps global consistency but costs VRAM.
• Advanced enhancer (use_advanced_enhancer): the "sharpen literally" stage. Adds adv_detail_boost (tanh-scaled emphasis on high-frequency detail), adv_high_pass_filter (edge emphasis), and adv_preserve_original (anchors the result to the raw embedding so cranking strength doesn't blow up). adv_attention_strength tunes its own attention mix, and adv_mlp_hidden_mult scales the MLP width.

The mental model from the original Capitan recipe: basic glues and stabilizes, advanced sharpens. They're meant to stack, which is why both default to on.

The inputs that matter, honestly:

• use_basic_enhancer / use_advanced_enhancer - the two master toggles. First thing to flip when you're A/B-testing whether the enhancement is even helping.
• basic_strength / adv_strength - start around 0.04–0.10. The author's own advice is start super low on the advanced stage; the ranges go negative to 2.0, and past about 0.3 you're asking for artifacts.
• adv_detail_boost - the one that visibly sharpens texture and edges. 1.9 default, up to 4.0. Great for fabric/stone, mean to skin.
• adv_preserve_original - your safety anchor. Higher means the output stays closer to the untouched embedding.
• basic_mlp_hidden_mult / adv_mlp_hidden_mult - the UI lets you type up to 200, but the code clamps these internally (12 for basic, 16 for advanced) and adds residual scaling, which is the pack's way of protecting you from yourself. Don't treat the max as a target.

There's no enhancement_profile here - the presets are gone, replaced by defaults that match capitan_daily's settings.

Outputs and wiring. Same as the siblings: positive CONDITIONING into ZImageTurboSampler, negative as a zeroed conditioning (inert at CFG 1 on this guidance-distilled model), and positive_prompt for saving. seed seeds the enhancement's random MLP init, so keep it fixed when you're comparing knob positions.

Install is pack-level: ComfyUI Manager → "ZImageTurboHQNodes", or git clone https://github.com/marduk191/ComfyUI-ZImageTurboHQNodes into ComfyUI/custom_nodes, restart. No pip deps; models come from Tongyi-MAI's Z-Image-Turbo HF repo.

The trap with this node is that it's genuinely fun to fiddle with and easy to over-tune. My advice: find a strength you like on one seed, then sit on it. The enhancement is a detail-retention boost, not a style switch - if you're spending an afternoon on the MLP multipliers, the fix you're looking for is probably a LoRA instead.