Before you start
- Source image (optional): Upload a sketch, collage, or image with Source image enabled; otherwise explicitly disable that stage to use the reference canvas.
- Editing instructions: Describe the finished image and identify what to preserve, replace, or reinterpret. Replace the retained living-room example.
Start with an upload and get one transformed image
Start with an uploaded sketch or collage and instructions describing the finished image. The named result is Transformed image (output:2), a single image preview—not a permanently saved PNG. The retained living-room prompt requests realistic windows and plants under warm afternoon sunlight; it is an example configuration, not evidence of an executed result.
The required preparation component is tea-time-labs/sweet-tea-nodes, linked here as a prerequisite, not an executable recipe. No component version or checksum is supplied. Execution mode, supported backends/platforms, GPU capacity, system memory, and comparable timing are not established. The 1024 × 1024 preparation workload is configuration metadata, not a measured performance profile.
The execution-effects inventory flags external data transfer and a paid ByteDance provider call at node 1. That assessment is partial and applicability is unverified; the other three nodes are uncharacterized. Treat this as a potential paid external operation, not local-only processing. The pack does not establish a price or an account/authentication setup procedure.
- Leave Source image enabled and upload the sketch, collage, or image to transform.
Look for: The upload supplies Painter's connected base image. LoadImage.image is declared in the native schema even though no value is saved in inputs; the widget filename interior_college.png does not establish an available asset.
- Replace Editing instructions with the desired finish and the elements to preserve, replace, or reinterpret. Choose Output size, then run.
Look for: The recipe is configured to return one Transformed image in the preview. Preservation language is a request, not a pixel guarantee.
Outputs: output:2
Follow the image path, not the mask names
LoadImage node 3, output slot 0 (IMAGE), feeds Painter node 7 input image. Painter output slot 0 (IMAGE) feeds ByteDanceSeedreamNodeV2 node 1 input model.images.image_1. Node 1 output slot 0 feeds SweetTeaPreviewImage node 2 input images, exposed as Transformed image (output:2). The terminal is a no-permanent-save preview transport; its historical SaveImageAdvanced metadata does not make it a current save node.
Both LoadImage and Painter declare MASK outputs, but neither MASK output is connected. Painter's string input mask holds retained drawing data; it is not an established protected-region mask sent to Seedream. This is reference-conditioned generation, not a demonstrated masked-region editing or original-layer compositing workflow. The word “collage” describes the input and intended transformation, not proof of a layer-preserving composite operation.
The saved model is seedream 5.0 pro. Its native schema also lists seedream 5.0 lite, seedream-4-5-251128, and seedream-4-0-250828, but the retained interface exposes no model selector and describes this recipe as Pro. The pack does not establish equivalent behavior or dynamic settings for switching models.
Prompt reasoning (thinking) is saved as true. Its native tooltip describes prompt optimization for better adherence and warns of substantially increased generation time, notably on Pro. It can only be disabled without reference images; both supported modes here provide one. The port contract labels thinking as text/prompt, whereas the native schema and interface identify a boolean. Use the native boolean type, not that port classification.
A label, subject, or supposedly protected area changes.
The image conditions generation; this graph does not composite an untouched original layer back into the result.
Check: Only Painter's IMAGE output reaches Seedream. No MASK output or ImageCompositeMasked node is connected.
Clarify the desired appearance in the prompt if reinterpretation is acceptable. If original pixels must remain exact, do not rely on this recipe as the preservation path. No suitable fallback recipe is retained.
Set generated dimensions, seed, and watermark
Output size defaults to (1K) 1024x1024 (1:1). The saved Custom width and Custom height are both 2048, but they take effect only when Output size is Custom. They do not override the selected 1K preset. Custom width accepts even values from 1024 to 3136; height accepts even values from 1024 to 2496. The full preset list and any additional size-combination limits are not supplied.
For a custom 2048 × 2048 request, use all three size controls below: select Custom and explicitly set both dimensions. Canvas dimensions do not replace these generated-size fields.
Seed has two retained representations: the interface default is -1, documented as requesting a fresh random seed per run, while the graph stores 902695217 and its native schema accepts integers from 0 to 2147483647. Treat -1 as an interface convention, not a valid native set-input value; its conversion is not shown. A nonnegative seed controls generation randomization but does not establish identical outputs across runs. The saved randomize widget is explicitly non-runtime, not a separate executable input.
Watermark defaults to false. Enabling it requests an AI-generated watermark on the resulting image; it does not change the reference path.
- Select custom output dimensions
- Makes generated dimensions use the current model.width and model.height values rather than the 1K preset.
- Set generated width to 2048
- Requests a width of 2048 pixels only when Output size is Custom.
- Set generated height to 2048
- Requests a height of 2048 pixels only when Output size is Custom.
- Use the retained nonnegative seed
- Sets the generation seed to 902695217 instead of relying on the interface's random-seed request; identical results are not established.
- Enable output watermark
- Requests an AI-generated watermark in the resulting image.
Changing Custom width or Custom height does not change the requested output size.
The preset determines whether custom dimensions are used.
Check: The retained preset is 1K even though both saved custom dimensions are 2048.
Select Custom, then set valid even dimensions in the generated-size fields, not the canvas fields.
Generation takes longer than expected.
Enabled prompt reasoning can increase generation time, but the pack does not establish the cause of any particular delay.
Check: Both supported modes send a reference image, so disabling reasoning is not supported here. No comparable timing is retained.
Keep reasoning enabled. Do not apply local VAE-tiling remedies to this graph: it contains no VAE decode node or tile-size input.
Use a blank canvas without an upload
The interface offers a Source image stage bypass (__bypass_3). This is a UI stage operation, not a native LoadImage input or a set-input adaptation. Merely leaving the upload empty is not the documented canvas-only setup.
With Source image disabled and drawing data empty, Painter supplies a blank reference using width 1024, height 1024, and background #000000 by default. Canvas width and height accept multiples of 64 from 64 to 4096. These settings describe the reference canvas, not the generated dimensions. Background color supplies the blank reference's color, not a guaranteed final-image background. Painter still sends an IMAGE reference to Seedream, so this is not reference-free text-to-image.
Drawing data maps to Painter input mask and defaults to an empty string. Its encoding is not established. This interface has no interactive drawing editor; use an externally prepared upload for a hand-drawn composition. The canvas controls below assume Source image has already been disabled; they do not remove its graph connection.
- Disable Source image using its stage bypass, then open Reference canvas. Leave Drawing data empty unless an integration supplies compatible data; choose canvas dimensions and background, enter Editing instructions, and run.
Look for: Painter supplies a blank reference image instead of an uploaded base image, and the recipe is configured to return one Transformed image.
Outputs: output:2
- Keep drawing data empty
- With Source image disabled, width, height, and background define the blank reference rather than undocumented drawing content.
- Use the retained black canvas background
- Supplies black as the blank reference background when Source image is disabled and drawing data is empty; it does not directly set final-image pixels.
There is no place to draw interactively.
The interface exposes canvas configuration and a retained data string, not a drawing editor.
Check: Painter's presence does not establish an exposed editor or documented drawing-data format.
Prepare the sketch elsewhere, enable Source image, and upload it. Do not invent a drawing-data encoding.
No upload is available, but Source image remains enabled.
The retained graph still connects LoadImage to Painter; an empty upload alone is not the documented switch to a blank canvas.
Check: Source image is optional in the interface but enabled by default, and the saved widget filename does not establish an available file.
Upload an image or explicitly disable the Source image stage before using the blank reference canvas.
Where this approach stops
- Reference-free text-to-image in this interface
- Interactive drawing editor
- Documented Painter drawing-data encoding
- Guaranteed original-pixel preservation
- Permanent PNG saving by this workflow
- Measured memory or run-time estimates
