Z-Image Turbo I2I (Replicate)

This is the img2img sibling of Replicate_prunaai_z_image_turbo, and it's the more interesting node of the pair for one reason: it can load LoRAs. In the cloud. The img2img variant of Alibaba's open Z-Image Turbo takes your image, your prompt, and optionally a list of LoRA weights via URL, and does the whole transformed-render on Replicate's hardware. If you've built a nice LoRA stack locally, this is the only node in this pack that lets you take it on the road without a GPU.

It's still an open model rented per call, so the same "why not run it locally" question applies - answer: no GPU. But the LoRA bit changes the math slightly: this node isn't just "generation without a card," it's "your tuned model without a card." That's a meaningfully different offer, and for a specific workflow it's worth the meter.

How it works

Standard pack engine around prunaai/z-image-turbo-img2img. Your input image goes up as PNG, the prompt rides along, and on the way the API loads any LoRAs you point it at. Output comes back as an IMAGE tensor. The img2img strength input controls how much of the original survives - same concept as local denoise strength, just with a cleaner name.

The inputs that matter

• prompt (required) - the text prompt for the transformation.
• image (required) - your input image.
• strength - default 0.6. 0.0 is no change, 1.0 is full transformation. For a gentle restyle stick around 0.4–0.6; for a big re-render push toward 1.0.
• num_inference_steps - default 8, and the tooltip adds a fun detail: it results in (num_inference_steps - 1) DiT forwards. Turbo rule stands - 8 is the target.
• guidance_scale - default 0, should stay 0 for Turbo. Same trap as the T2I node.
• lora_weights - the headline feature. A URL list as a string, e.g. ['https://example.com/lora1.safetensors', 'https://example.com/lora2.safetensors']. Supports .safetensors, tar, and zip URLs. The files have to be fetchable - Replicate's server downloads them, so no local paths.
• lora_scales - one float per LoRA, defaults to 1.0 if omitted. Must match the count in lora_weights.
• seed - set for reproducibility.
• output_format / output_quality - jpg/80 by default.

Single output: image, an IMAGE tensor.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", restart. Only dependency: replicate>=1.0.7. Auth as an environment variable:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• LoRAs must be URL-addressable. Local paths won't work - Replicate's workers need to fetch the file themselves. Host it somewhere public or on the Files API first.
• Mismatched LoRA/scale lists. The tooltip is blunt: scales must match the number of weights. A mismatch is an error, not a graceful default.
• Strength surprises. 1.0 is a full transformation; expect the source to be unrecognizable. If you want "same person, new background," that's a much lower strength than you think.