Virtual Try-On Candidates (Mock)

It says "Mock" right in the title, and that's the whole story. This node generates on-model candidate images by blending pixels, not by inference. There is no diffusion model, no FASHN API call, no GPU work. It's a placeholder that lets you exercise the entire pipeline - candidate generation, evaluation, ranking, export - without spending a cent. The real virtual try-on (FASHN) is milestone 3 and isn't shipped yet, so this is what you get, on purpose.

The mechanism is refreshingly transparent. It takes your product image, resizes it to the model reference's dimensions with bilinear interpolation, then alpha-blends the product over the model. Each candidate uses a slightly stronger blend (alpha starts around 0.24 and creeps up per candidate) plus a small brightness offset, all seeded and deterministic. Same seed, same candidates, forever. The seed also gets recorded per candidate, so provenance is intact even though nothing generative happened.

Inputs worth knowing:

• candidate_count (default 3, range 1–4) - how many on-model variants. The 4-cap is enforced twice, here in the node and again by AIFS_MAX_CANDIDATES in the environment.
• seed (default 42) - determinism control. Bump it to get a different blend pattern; there's no aesthetic meaning to it.
• product_front, model_reference, plus product_bundle and product_specification - the images feed the blend, the bundle and spec ride along so the downstream manifest keeps the full provenance trail.

Outputs: candidate_images is a batched IMAGE (all candidates stacked - the node previews them as a strip), candidate_set is the AFS_CANDIDATE_SET object carrying per-candidate IDs, seeds, and latency, and candidate_metadata_json is the same thing readable. Wire candidate_set into the evaluation stage - that's the handoff that matters.

The honest take, and it matters: the blended results are not a virtual try-on. You will not judge garment quality from them, because there is no garment rendering happening - the product is literally composited onto the model like a cheap collage. The pack's own README says the mock "verifies plumbing and UX only; it does not predict the visual quality of real virtual try-on." Treat this node as orchestration rehearsal. If what you actually want is a real VTO today, this pack isn't that yet - you'd be looking at local approaches or a different provider integration.

Setup is the pack standard: Manager search "AI Fashion Studio", or clone into custom_nodes and restart. No model downloads, no API key, no environment variables needed for the mock slice. It lives under AI Fashion Studio / Mock MVP.

Where people get tripped up: they wire only the two images and forget the product_bundle and product_specification sockets - those are required, not optional, and the graph won't run without them. And if you bump candidate_count past 4, the node raises rather than quietly trimming. That's the ceiling working as intended.