Image upscale and refinement: mechanisms and practical controls

Produce one image named Refined and upscaled image, exposed at output:24. Begin with an uploaded illustration, the preset prompts, 2× enlargement, initial denoise 0.35 and tile denoise 0.3. Add a description of your subject to the positive prompt when you want more specific guidance.

This recipe enlarges an existing image through generative refinement. Initial diffusion, enlarged tile redraw and seam correction can all change content. Lower strengths retain more incoming structure without establishing unchanged pixels. The Resource task family is upscale; resolution increase alone does not establish restoration, identity preservation or a detector-based detailer.

The final output comes from node 12, slot 0, through node 24 (SweetTeaPreviewImage) input images. The interface describes preview transport without permanent saving. Intermediate images are not exposed as named outputs.

The graph names three fixed files: DasiwaIllustriousAnime_alluringabyss.safetensors at node 2 (CheckpointLoaderSimple, ckpt_name), AddMicroDetails_Illustrious_v5.safetensors at node 4 (LoraLoader, lora_name) and 2x-AnimeSharpV4_RCAN.safetensors at node 10 (UpscaleModelLoader, model_name). The retained interface exposes strength controls, not file selectors. These presets target illustrated imagery; compatibility or quality with replacement files is not established here.

The preparation inventory also lists comfyui_ultimatesdupscale and tea-time-labs/sweet-tea-nodes as required node packs. Graph metadata records comfy-core 0.16.4 and UltimateSDUpscale 1.7.0. Those metadata values do not establish a complete compatible environment. All five preparation component records are marked retired and lack artifact checksums and provider versions, so current availability and exact artifact identity remain unresolved.

No supported backend or platform, GPU capacity, RAM capacity or comparable duration is retained. The preparation record's 1024×1024 standard workload is a reference configuration, not an input-size restriction or a measured execution profile.

Node 22's IMAGE output, slot 0, enters node 7 (VAEEncode) through pixels; node 2 supplies its VAE through slot 2. Node 8 (KSampler) receives node 7's latent through latent_image and denoises it using node 4's LoRA-modified MODEL. Node 9 (VAEDecode) receives node 8's LATENT output through samples and converts it back to an image using the checkpoint VAE. This initial image-to-image pass precedes enlargement.

Node 11 (UltimateSDUpscale) receives node 9's image, node 10's UPSCALE_MODEL, the same modified diffusion model, the same VAE and both conditioning streams. It enlarges the image and performs prompt-guided tile redraw followed by the configured seam correction. Node 12 (ImageSharpen) performs final sharpening before node 24 receives the image. No detector or separate region-detailer node is present.

The conditioning branches differ. Node 3 (CLIPSetLastLayer) is configured with stop_at_clip_layer=-2. Its CLIP output enters node 4, whose modified CLIP supplies positive-text encoding at node 5 (CLIPTextEncode). Negative-text encoding at node 6 uses node 3 directly. Both node 8 and node 11 receive node 5's positive and node 6's negative conditioning. LoRA model strength therefore affects both diffusion stages, while LoRA CLIP strength affects their shared positive encoding only.

Node 22's MASK output is unconnected, and there is no original-layer composite or alpha-joining path. The retained interface explicitly states that uploaded transparency is not preserved. The compiled operation has no rule bindings or established output properties; these explanations follow the retained wiring and control descriptions.

Initial denoise (initial_denoise → node 8 denoise) defaults to 0.35; tile denoise (tile_denoise → node 11 denoise) defaults to 0.3. Both accept 0–1 in increments of 0.01. The first controls refinement before enlargement; the second controls enlarged tile redraw. Lower values retain more incoming structure, while higher values permit stronger reinterpretation. Neither controls the separate seam-correction strength.

The positive preset is addmicrodetails, masterpiece, best quality, very aesthetic, absurdres, highres, fine detail, sharp focus, intricate linework, crisp edges, detailed shading. The negative preset is worst quality, low quality, lowres, blurry, jpeg artifacts, compression artifacts, noise, grain, ugly, deformed, bad anatomy, watermark, signature, text. Their ports target node 5 text and node 6 text. Both diffusion stages share these prompts. Negative terms discourage traits; they do not guarantee removal.

Each stage has independent sampling controls:

Interface port	Node/input	Default	Meaning
initial_steps	8 / steps	20	Initial denoising steps
initial_guidance	8 / cfg	9	Initial prompt guidance
initial_sampler	8 / sampler_name	dpmpp_2m	Initial sampling algorithm
initial_scheduler	8 / scheduler	karras	Initial noise schedule
tile_steps	11 / steps	20	Steps for each redraw tile
tile_guidance	11 / cfg	8	Tile prompt guidance
tile_sampler	11 / sampler_name	dpmpp_2m	Tile sampling algorithm
tile_scheduler	11 / scheduler

At node 11, upscale_factor targets upscale_by, default 2, with a range of 0.05–4 in increments of 0.05. It multiplies width and height; values below 1 reduce dimensions. tile_width and tile_height each default to 512 pixels and describe processing regions, not final dimensions. Each must be at least 64 and use multiples of 8. Larger tiles require more memory.

tile_pattern → mode_type defaults to Chess and sets redraw order. tile_mask_blur → mask_blur defaults to 8 pixels, accepts integer values 0–64 and softens boundaries. tile_padding defaults to 32 pixels, accepts nonnegative multiples of 8 and adds surrounding context and memory cost. The port inventory labels tile mask blur as a mask, but the interface and native schema define an integer radius, not an uploaded mask.

tile_batch_size → batch_size defaults to 1 and accepts 1–4096. Larger batches can reduce processing time but use more memory and yield different results; they affect only the main redraw, not seam correction. force_uniform_tiles defaults to true, keeping boundary tiles at the configured dimensions. Turning it off allows smaller boundary tiles.

tiled_decode defaults to true and decodes node 11's tile outputs in smaller regions. It does not modify node 9, which remains a separate VAEDecode. This recipe exposes no decoder tile_size input, so the domain's generic smaller-decode-tiles adaptation cannot be bound here. No numerical memory saving is established.

Node 11's seam_fix_mode defaults to Half Tile + Intersections. Correction adds generative processing after tile redraw. seam_fix_denoise defaults to 1, its maximum, so corrected regions can change substantially even with tile denoise at 0.3. It accepts 0–1 in increments of 0.01. Seam mask blur defaults to 8 pixels and accepts integer values 0–64. Seam padding defaults to 16 pixels and uses nonnegative multiples of 8. Both target their identically named inputs at node 11.

seam_fix_width defaults to 64 pixels and uses nonnegative multiples of 8, but applies only in Band Pass mode. It is not the width control for the default correction mode. Disabling correction skips its extra processing. The pack does not enumerate the exact disabled option string, so use the runtime's offered disabled option rather than constructing a value. The lower-strength control below is a proposed trial, not an established optimum.

Node 12's advanced controls are sharpen_amount → alpha (default 0.3, range 0–5, increment 0.01), sharpen_radius (default 1 pixel, integer range 1–31) and sharpen_sigma → sigma (default 0.8, range 0.1–10, increment 0.01). Radius changes the edge neighborhood and sigma changes the Gaussian spread. Sharpening is the only exposed bypassable stage. Bypassing it yields the unsharpened upscale, still subject to all preceding generative changes. The contract has no bypass delta, so bypass remains a plain interface instruction.