IAMCCS Qwen Multi-Gen

IAMCCS_QwenMultiGen runs newline-separated prompt variants through a single Qwen image-edit execution loop: one source image in, a whole list of edits out, all stacked in one batch. It exists because dataset-creation workflows - the pack's README positions it as a companion to IAMCCS QE Prompt Enhancer - routinely need "the same image, ten different edits" without duplicating the sampling chain ten times. This node is that loop.

Qwen-Image-Edit is the open image-editing model that ruled 2025 before Flux Klein ate into its territory; the community read is that Klein is often better at pure edits while Qwen-Edit is cheaper to run and less finicky about character sheets. For batch iteration, though, a loop node is a loop node - you're not choosing between models here, you're automating the one you already loaded.

How it works

Feed it a loaded Qwen model, CLIP, VAE, and your source image. The multi_prompt field holds one edit per line; it splits on separator (newline default). For each line it encodes the Qwen conditioning (image + prompt), applies the reference_latents_method (default index_timestep_zero) to bind the edit to the source, samples with seed + index, decodes, and saves a PNG to ComfyUI's output folder using output_prefix plus a prompt slug. save_images defaults to true here - flip it off if you just want the tensor back.

The defaults tell you what kind of runs this expects: 4 steps, CFG 1, denoise 1.0. That's the fast distilled-edit profile - it's built for iterating dozens of variants quickly, not for fine quality on a single hero image.

Inputs and outputs

Required: model, clip, vae, image, multi_prompt, seed, steps, cfg, sampler_name, scheduler, denoise. Optional: separator, reference_latents_method, output_prefix, save_images. Outputs: images (a batch of one image per prompt line) and count (how many ran).

That count output is quietly useful: wire it into anything that needs to know the batch size downstream, or just use it to sanity-check that all your lines actually parsed. An empty multi_prompt returns a blank image and count 0 rather than erroring.

Installing it

Part of the IAMCCS-nodes pack. ComfyUI Manager → search IAMCCS, or:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Restart ComfyUI; it sits under IAMCCS/Qwen. No pip requirements ship with the pack. The real prerequisite is the model: you load a Qwen-Image-Edit checkpoint, its text encoder, and VAE yourself, then hand them to this node.

Where people get burned

• Forgetting it's a loop driver, not a loader. Connect real model/CLIP/VAE or you'll get an immediate failure.
• save_images on by default can quietly fill your output folder with dozens of PNGs on a long variant list. If you only want the batch output, turn it off.
• Expecting seed stability across runs: seeds increment per line (seed + index), which is deliberate - re-running with the same seed gives you the same batch, and bumping the seed shifts all variants.
• Qwen-Image-Edit has its own licensing terms to check if you're building a commercial dataset with it; that's a model-policy question, not a node question, but worth knowing before you scale up a batch run.