FLUX virtual try-on: mechanisms and practical controls

Produce Try-on result by starting with a clear person photo and one garment reference. Show the body or clothing area you want dressed. For several pieces, prepare one combined reference and identify the pieces in Styling instructions; this recipe accepts one garment image.

The selected output is the image preview at output:4. Its cardinality is determined at runtime, so the retained contract does not promise an exact image count. The workflow does not automatically save an image file. The preparation record's 1024 × 1024 workload is descriptive metadata: no width or height control appears in the graph, and that record does not establish the dimensions of every returned image.

Node 1 (LoadImage) sends output slot 0 to 7.person; node 5 (LoadImage) sends output slot 0 to 7.garment. Node 7 (FluxVTONode) uses these images with prompt and seed for generation through BFL. Its IMAGE output at slot 0 feeds 4.images. Node 4 is SweetTeaPreviewImage, titled Try-on result. These are existing graph connections; change the uploads at the two LoadImage nodes rather than replacing 7.person or 7.garment with scalar settings.

The native description states that generation sends the input images and instructions to BFL and incurs provider-priced usage. The retained effects assessment also identifies external transfer and a paid provider call, but is partial and marks applicability unverified. It supplies no exact price or complete characterization of every node.

This is reference-guided image generation. The garment port's identity_reference role names how the image is used; it does not establish identity preservation. Similarly, asking for upper-body changes does not establish masked regional editing. The graph exposes no mask or final original-layer composite, and the provider's internal generation stages are not retained. Face, pose, garment-detail and unchanged-pixel preservation therefore remain requests to assess in the preview, rather than established output properties.

The retained preparation inventory lists tea-time-labs/sweet-tea-nodes as a required node pack, linked below as a prerequisite resource. No version, commit or checksum is pinned. This resource is a dependency, not an alternative executable recipe.

The graph must have its LoadImage, FluxVTONode and SweetTeaPreviewImage classes available. The inventory does not establish a compatible ComfyUI version or complete installation procedure for those classes. Before submitting images, establish that the environment can execute this recipe's BFL generation path. Credential, account and billing configuration details are absent from the pack; no particular credential name or setup command can be inferred here.

Execution mode, backend and platform are unspecified. There is no scoped GPU-capacity recommendation, system-memory recommendation or comparable run timing. The listed node pack and workload metadata alone do not establish that a particular machine or provider configuration is ready.

Styling instructions maps to 7.prompt. Describe the garment, fit, placement or intended area, and identify individual pieces when using a combined reference. The saved default requests full-outfit replacement. The field is required in the interface and schema despite its tooltip calling the instruction optional; retain a meaningful instruction because blank or omitted behavior is not established.

Seed maps to 7.seed, described by the schema as the seed used to create noise. The interface default is -1 and its note describes that value as requesting a new random seed on each run. However, both the interface constraints and node schema have minimum 0, and the saved graph input is 0. The saved widget's randomize entry is marked non-runtime; it does not establish how the interface translates -1 into a valid runtime seed. A fixed nonnegative integer avoids that ambiguity. The retained maximum is 18446744073709552000; seed 0 is the saved value and fits the stated range.

A fixed seed controls the noise-seed input, without establishing identical results across runs or provider changes. To compare seed choices, hold both images and the prompt unchanged. To assess a wording change, hold both images and the fixed seed unchanged. The prompt adaptation below is an editable request example, with no claimed executed result or enforced spatial boundary.

After supplying the images and configuring the request, run the recipe in the configured environment. Successful generation returns Try-on result through output:4. Assess clothing placement, fit, garment details and any face or pose details that matter to the task. Each new generation uses provider-priced usage.

Choose a remedy according to the visible problem. This graph has no independent mask, preservation-strength control or local decode stage to adjust. If a run fails, use the actual error to identify whether image loading, node availability or provider configuration needs attention; the retained pack supplies no error-specific provider recovery procedure.