SeedVR2 image upscale: mechanisms and practical controls

Start with one source image and the model pair listed under prerequisites. The default resize multiplies width and height by 4, giving approximately 16 times the source pixel count before restoration. This is not merely interpolation: SeedVR2 subsequently processes the resized image.

The named output is Restored image, port output:9: node 73 output slot 0 feeds SweetTeaPreviewImage node 9 through images. There is one image output, not separate original and restored outputs. The retained terminal does not establish a saved filename or file format.

Inspect fine text, markings, edges and color in the result. Restoration can reinterpret details rather than recover exact missing information. The saved watch-image filename is only a graph selection; no executed example or measured result is supplied.

This guide covers recipe version 0.1.6 and the selected graph revision, not interchangeable SeedVR2 workflows. The graph selects seedvr2_3b_int8_convrot.safetensors through UNETLoader node 76, with weight_dtype=default, and seedvr2_ema_vae_fp16.safetensors through VAELoader node 82. Their declared model directories are diffusion_models and vae, respectively. Both files are required. Preparation identifies provider version main, but supplies no model artifact checksums; that version label alone is not an immutable model pin. The filenames do not establish the precision of every runtime operation.

The required node pack is tea-time-labs/sweet-tea-nodes, including SweetTeaPreviewImage. The runtime must also provide the graph's core classes: LoadImage, JoinImageWithAlpha, ResizeImageMaskNode, SeedVR2Preprocess, VAEEncodeTiled, SeedVR2Conditioning, KSampler, VAEDecodeTiled, SeedVR2PostProcessing, UNETLoader and VAELoader. Core node metadata records version 0.27.0; this is retained version evidence, not a tested minimum-version claim. No node-pack commit, backend list or platform compatibility matrix is supplied.

Preparation's Standard configuration records 1024×1024, without frames, batch, precision or memory mode. It is not a measured run or a guaranteed input/output size. GPU capacity, system RAM and duration remain unknown. The linked prerequisite resources are dependencies, not executable fallback recipes.

LoadImage node 1 sends IMAGE slot 0 and MASK slot 1 to JoinImageWithAlpha node 81, inputs image and alpha. The retained join implementation resizes the mask if needed, computes alpha as one minus that mask, and concatenates it with the input's first three RGB channels. RGB preservation applies at this join stage only. It does not establish final transparency or background removal.

The retained operation metadata classifies this recipe as remove_background, but the actual graph has no background-removal model or foreground-mask inference node. Joining the loader's existing mask is not subject segmentation. This guide therefore uses the supported task families upscale and restore, rather than treating the classification as proof of background removal.

Node 81 slot 0 feeds ResizeImageMaskNode node 80 at input. Its saved mode is scale by multiplier. Node 80 slot 0 branches to SeedVR2Preprocess node 79 at resized_images and to postprocessing node 73 at original_resized_images. The latter branch supplies the resized-source reference, not an original-pixel bypass to the terminal. The graph establishes preprocessing's position, but not its internal transformation.

Node 79 slot 0 feeds tiled encoder 78 at pixels. VAE node 82 slot 0 supplies vae to both encoder 78 and decoder 74. Encoder 78 slot 0 supplies both SeedVR2Conditioning node 77 at vae_conditioning and sampler 75 at latent_image. Model node 76 slot 0 supplies model to nodes 77 and 75. Conditioning node 77 slots 0 and 1 supply the sampler's positive and negative inputs: there is no text-prompt encoder in this path. The retained graph establishes these connections, not the internal construction of the two conditioning outputs.

KSampler node 75 slot 0 feeds decoder 74 at samples; decoder slot 0 feeds postprocessor 73 at images. Postprocessing then supplies the named output. Resizing, latent processing and decoding mean that a recognizable result is not proof that original pixels, text or tiny markings remain unchanged.

Scale factor maps to node 80 input resize_type.multiplier: default 4, range 0.01–8 in 0.01 increments. Above 1 enlarges, 1 keeps source dimensions and below 1 reduces; restoration runs in all three cases. The multiplier applies while resize_type remains scale by multiplier, as saved in this revision. Other resize modes are listed by the schema, but their required parameter sets are not retained here. Larger dimensions increase the image data processed; a smaller scale is not a quantified memory remedy.

Resampling method maps to node 80 input scale_method, default lanczos. Available values are nearest-exact, bilinear, area, bicubic and lanczos. Lanczos suits enlargement, area reduction, and nearest-exact hard pixel boundaries during the initial resize only; subsequent restoration can change those boundaries.

Color matching maps to node 73 input color_correction_method, default none. According to the retained schema, none skips color transfer but retains geometry alignment; lab transfers color in CIELAB space; wavelet transfers low-frequency color while retaining restored high-frequency detail; adain matches per-channel mean and standard deviation. Color transfer uses node 80's resized source as reference, not a separate uploaded reference. No mode is a pixel-preservation switch. Comparative quality or speed measurements are not supplied.

The advanced controls bind to KSampler node 75: Denoise → denoise (default 1, range 0–1, step 0.01); Steps → steps (default 1, range 1–10000); Conditioning guidance → cfg (default 1, range 0–100, step 0.1); Sampler → sampler_name (default euler); Scheduler → scheduler (default simple). Sampler and scheduler choices are runtime-provided; no complete option list is retained.

Denoise controls the fraction of the denoising schedule used; steps controls sampling iterations; guidance applies to image-derived conditioning, not a user text prompt. Retain the baseline initially. Lower denoise or changes to steps, guidance, sampler and scheduler are experimental for this one-step setup, not established quality improvements. The generic sampler tooltip about lower denoise maintaining structure is not evidence of exact preservation or a tested SeedVR2 tuning recipe.

Seed controls sampling noise. The saved native graph uses 959948902156062; the interface instead declares default -1 for randomization, while both its numeric constraint and the native schema require nonnegative values. The native schema lists a maximum of 18446744073709552000. Treat -1 as an advertised UI convention whose translation is not established here, not a valid native seed delta. If the interface rejects it, use a fixed nonnegative seed rather than writing -1 directly into KSampler. A fixed seed keeps sampling noise consistent for comparisons, without guaranteeing identical output across runtimes.

For useful comparisons, keep the source, scale and sampling baseline fixed and change one color or tiling setting at a time. There are no retained executed comparisons that establish a best alternative.

Input tiling belongs to VAEEncodeTiled node 78; output tiling belongs to VAEDecodeTiled node 74. Both save tile_size=512 and overlap=128, measured in pixels. Encode tile size permits 64–4096 in steps of 64; decode permits 64–4096 in steps of 32. Both overlap inputs permit 0–4096 in steps of 32, but keep overlap below the corresponding tile size. These controls are independent; changing output tiling does not change input tiling.

The retained decoder implementation converts tile dimensions by VAE compression factors and calls tiled decoding. If tile size is less than four times overlap, it reduces effective overlap to tile size divided by four. Thus a 256-pixel decode tile with saved overlap 128 uses effective overlap 64. This is an implementation-derived example, not a recorded run. It also explains why the displayed overlap can differ from the effective value. No corresponding encode-overlap adjustment rule is supplied.

Smaller spatial tiles can reduce memory requirements at their respective VAE stage, but can change seams and processing work. No quantitative memory or speed saving is established. Decode tiling does not change model loading or denoising memory requirements.

Both encoder and decoder also save temporal_size=4096 and temporal_overlap=8. Their schema says these are only used for video VAEs; they are not still-image spatial-tiling controls.