Amazing Upscale Node Workflow for DIT Model

This workflow is designed for simple logic amazing upscale nodes in the DIT model. It supports common applications for Flux, Hunyuan, and SD3. The workflow tiles the initial image into smaller pieces, uses an image-interrogator to extract prompts for each tile, and performs an accurate upscale process. This approach minimizes hallucinations and ensures proper condition handling.

We hope you enjoy using it!

What's New

TeaCache Sampler Integration for Hunyuan Video

Thanks to the contributions from the TeaCache code repository (ali-vilab/TeaCache) and code references from facok/ComfyUI-TeaCacheHunyuanVideo, we’ve added support for the TeaCache sampler.

• How to Use: Replace the samplercustomadvanced node in the official workflow with the TeaCache sampler node. Adjust the acceleration rate as needed to start using it.

• Performance: In testing with an NVIDIA 4090, rendering a 720×480 resolution video with 65 frames took only 55 seconds using a speedup factor of x2.1. This is approximately twice as fast as the original method.

• Caution: While the TeaCache sampler significantly accelerates processing, it may reduce image quality and dynamic effects. Use with discretion.

• Precision Support: Supports bf16 and fp8.

https://github.com/user-attachments/assets/af06b9d3-9c84-4a83-ba90-eb4ec4bb2e99

---

Instructions

Smart Tile Workflow

Smart Tile is an object-aware tiled img2img workflow for ComfyUI. Instead of cutting every image into equal squares, it lets you build variable-size tiles around visual structure: face, eyes, hands, text, subject, clothing, foreground details, and background/context regions.

Smart Tile is designed for workflows where small areas need more detail than the full image. You can draw or auto-detect tiles, process one tile at a time through your sampler, then assemble the final image with feathering, masks, semantic priority, color correction, and optional pixel alignment.

What It Solves

• Avoids cutting important objects through the middle when a simple grid would split a face, hand, or text.
• Allows different tile sizes, so a face tile can be small and high detail while the background stays large.
• Sends one tile at a time through img2img, so the number of tiles can be 4, 8, 16, or anything produced by the editor.
• Keeps high-detail focus tiles visible during assembly with priority, layer, and semantic ranking.
• Supports mask-aware pasteback, soft detail overlay, final-only assembly, and black-base canvas mode.

Visual Tile Editor

TTP Smart Tile Interactive Crop is the recommended starting point. It can load an image like ComfyUI's official Load Image, or receive a connected source_image. The editor stores the current layout in a hidden layout_json widget, so your workflow keeps the tile plan.

Useful editor actions:

• Replace grid: replace the whole layout with an even grid.
• Grid in T#: subdivide the selected tile without rebuilding the whole layout.
• Mask to Tile: turn painted regions into object tiles.
• Refresh masks: after manually moving sub-tiles from a masked tile, re-crop the inherited masks to the current tile boxes.
• Fill gaps: add background tiles for uncovered areas.
• Auto Tile: run SAM3.1 or QwenVL3 detection and write the detected layout back into the editor.

When subdividing a masked tile with Grid in, the Mask mode can crop the original object mask into child tiles or skip empty mask children. The child tiles keep a parent mask source, so Refresh masks can update them after manual edits.

Core Nodes

Node	Purpose
TTP Smart Tile Interactive Crop	Load/connect an image, create manual/painted/auto tiles, and output a variable-size tile_set.
TTP Smart Tile Set Preview	Preview a tile set as a contact sheet or one selected tile.
TTP QwenVL3 Local Loader	Load a local QwenVL tagging model from ComfyUI/models/text_encoders.
TTP Smart Tile QwenVL Prompt Set Builder	Generate per-tile prompts before loop processing.
TTP Smart Tile Semantic Rank	Classify tiles and write recommended layer, priority, scale weight, and composite metadata.
TTP Smart Tile Loop Source	Output one tile at a time for VAE Encode / sampler / VAE Decode.
TTP Smart Tile Loop Collect	Collect each processed tile back into the same tile set.
TTP Smart Tile Image Upscale Prep	Optionally upscale or resize each tile before sampling, with a megapixel cap.
TTP Smart Tile Output Size Estimate	Estimate final output scale/resolution from processed tile sizes.
TTP Smart Tile Assemble	Paste processed tiles back with feathering, masks, color correction, priority, optional GPU paste, and optional alignment.
TTP Smart Tile Save Final Image	Save only the final completed loop result and embed workflow metadata.

Recommended Loop Workflow

TTP Smart Tile Interactive Crop
  -> TTP Smart Tile QwenVL Prompt Set Builder (optional)
  -> TTP Smart Tile Semantic Rank (optional)
  -> TTP Smart Tile Loop Source
  -> VAE Encode / Sampler / VAE Decode
  -> TTP Smart Tile Loop Collect
  -> TTP Smart Tile Semantic Rank (optional final refresh)
  -> TTP Smart Tile Output Size Estimate (optional)
  -> TTP Smart Tile Assemble
  -> TTP Smart Tile Save Final Image

Use TTP Smart Tile Loop Source and TTP Smart Tile Loop Collect instead of manually duplicating sampler chains. The loop source outputs the current tile image and the current tile prompt. After the sampler finishes, loop collect stores the result and advances to the next tile.

By default, TTP Smart Tile Assemble uses assemble_mode=final_only. Connect TTP Smart Tile Loop Collect.done to TTP Smart Tile Assemble.done so unfinished loop runs return a lightweight preview while done=false, then perform the full assemble once after the last tile. Switch assemble_mode to always only when you really want a full recomposite after every tile. If pixel alignment is enabled, unfinished loop runs are automatically treated as final-only to avoid repeated expensive alignment passes.

Complete Example Workflow / 完整范例工作流

Example workflow / 示例工作流: examples/smart_tile_qwen_sam_loop_example_v2.json

这份工作流展示了当前推荐的 Smart Tile 完整链路：先用 TTP Smart Tile Interactive Crop 和 sam3.1 做 Auto Tile 分块，再用 TTP Smart Tile QwenVL Prompt Set Builder 生成每块提示词。随后 Semantic Rank 写入语义优先级，Loop Source 把 tile 一张一张送入 img2img，Image Upscale Prep 负责采样前放大，Loop Collect 收集结果，最后由 Output Size Estimate 估算画布尺寸，Assemble 在 final_only 模式下一次拼合，Save Final Image 只保存最终图。

This workflow is the recommended full Smart Tile loop. It uses TTP Smart Tile Interactive Crop with sam3.1 Auto Tile, prepares per-tile prompts with TTP Smart Tile QwenVL Prompt Set Builder, ranks tiles with Semantic Rank, sends one tile at a time through img2img with Loop Source, upscales each tile before sampling with Image Upscale Prep, collects results with Loop Collect, estimates the final canvas with Output Size Estimate, assembles once in final_only mode, and saves only the completed result.

Workflow overview / 工作流全景

Key views / 关键视图

Interactive editor / 交互分块编辑器	Tile set preview / 分块预览
Weight preview / 权重预览	Final preview / 最终预览

Key settings / 关键设置

Node / 节点	Example setting / 示例设置	Notes / 说明
Interactive Crop	auto_detect_mode=sam3.1, default_pad=32, default_blend=32, auto_object_padding=64, auto_max_tiles=16	Use Auto Tile to create an object-aware layout, then set auto_detect_mode=none if you want to freeze manual edits. / 用 Auto Tile 生成语义分块；如果之后要固定手动编辑结果，可以把 auto_detect_mode 改回 none。
QwenVL3 Local Loader	qwen3vl_4b_fp8_scaled.safetensors	Replace this with the QwenVL model installed in ComfyUI/models/text_encoders. / 按你本机 models/text_encoders 里的 QwenVL 模型替换。
QwenVL Prompt Set Builder	reference_image_mode=first_message, prompt_preset=tile_img2img_prompt, output_language=chinese	Builds prompts before the tile loop, then Loop Source.prompt feeds the text encoder. / 在循环前一次性生成每块提示词，然后由 Loop Source.prompt 送入文本编码。
Image Upscale Prep	scale=2.5, round_to=8, max_megapixels=1.5, use_upscale_model=true	Enlarges each tile before img2img while keeping each tile under the megapixel cap. / 采样前放大每块 tile，同时用百万像素上限控制显存。
Loop Source / Loop Collect	Process All Tiles workflow	The tile count can be 4, 8, 16, or any layout count; no manual sampler duplication is needed. / 分块数量可以变化，不需要手动复制多套 sampler。
Output Size Estimate	focus_weighted	Estimates the final canvas from processed tile sizes, giving more influence to focus/detail tiles. / 根据处理后的 tile 尺寸估算最终画布，并更重视细节块。
Assemble	assemble_mode=final_only, mask_blend_mode=mask_feather, color_correction=histogram, small_tile_on_top=true, auto_composite_policy=safe_auto	Assembles only after the loop is done, uses mask-aware feathering, and keeps small detail tiles above larger context tiles. / 循环结束后只拼合一次，使用蒙版羽化，并让小细节块优先于大背景块。

Before running the example, replace model filenames such as the diffusion model, CLIP/text encoder, VAE, LoRA, and upscale model with files installed on your machine. The workflow is meant as a wiring reference; model choices can be changed freely.

运行示例前，请把 diffusion model、CLIP/text encoder、VAE、LoRA、upscale model 等模型名替换成你本机已经安装的文件。这份工作流主要作为连线参考，具体模型可以自由替换。

Auto Tile: SAM3.1 or QwenVL3

auto_detect_mode controls how Auto Tile analyzes the image:

Mode	Required input	Notes
none	none	Auto Tile is disabled; use manual grid, painted masks, or saved layout.
sam3.1	official SAM3/SAM3.1 model to vision_model, plus CLIP or vision_conditioning	Uses ComfyUI's official SAM3 Detect path. The built-in auto_prompt is encoded internally when clip is connected.
qwenvl3	TTP QwenVL3 Local Loader -> qwen_vl_model	Uses QwenVL for bbox JSON. Model files are selected from ComfyUI/models/text_encoders. If Qwen returns one large full-frame box, Smart Tile splits it into useful detail tiles.

For QwenVL3, TTP Smart Tile Interactive Crop does not read .safetensors directly. Add TTP QwenVL3 Local Loader, choose the QwenVL model file, then connect its qwen_vl_model output into Interactive Crop.

auto_max_tiles limits the total Auto Tile layout after automatic gap filling, so detected objects and generated background gap tiles stay within the same cap.

Per-Tile Prompts

For manual prompt workflows, connect TTP Smart Tile Loop Source.prompt to your text encoder path. For automatic prompt workflows, place TTP Smart Tile QwenVL Prompt Set Builder before the loop:

Interactive Crop
  -> QwenVL Prompt Set Builder
  -> Semantic Rank
  -> Loop Source

The prompt builder can use a full image, every tile, or a contact sheet as QwenVL visual context. It caches results by model file, tile hash, prompts, and seed, so reruns do not need to interrogate unchanged tiles again.

Upscale Prep and Size Estimate

TTP Smart Tile Image Upscale Prep prepares each loop tile before img2img sampling. It can use a connected ComfyUI UPSCALE_MODEL through the same tiled upscale-model path as the built-in upscale node, or fall back to lanczos, bicubic, bilinear, area, or nearest resize when no model is connected or use_upscale_model is off. scale sets the requested enlargement, max_megapixels caps the final tile pixel count, and round_to snaps the final width/height after the cap. When the cap is active, the node rounds down so the rounded tile stays under the megapixel budget. Tile coordinates are not changed; they remain in original-image space so assemble can map the processed tile back by sample_box and output_scale.

TTP Smart Tile Output Size Estimate reads the processed tile_set after Loop Collect and reports output_scale, final width/height, separate scale_x/scale_y, and a per-tile info log. The default median strategy matches Assemble's automatic tile-scale inference, and the output_scale output can be connected directly to TTP Smart Tile Assemble.output_scale. focus_weighted uses semantic scale weights so low-detail full/background tiles do not drag the final-only canvas scale below high-detail face/eye/text tiles. For final-only loops, connect TTP Smart Tile Loop Collect.done to both this node's done input and TTP Smart Tile Assemble.done; while done=false, this node returns a deferred zero-scale placeholder and skips tile scanning, then estimates once when done=true. Mixed tile scales are reported in the info string so capped or unevenly enlarged tiles are visible before final assembly.

Assembly, Masks, and Layer Priority

TTP Smart Tile Semantic Rank is optional, but useful after QwenVL prompting or auto/manual tile creation. It classifies each tile as background, subject, face, eyes, hands, text, detail, or normal from existing labels/captions/prompts, then writes semantic score, scale weight, recommended layer, priority, occlusion priority, and composite mode metadata back into the tile set. With apply_composite_rank on, face/eyes/text/detail tiles are promoted above background/context tiles and are marked for soft overlay blending.

Assembly options worth starting with:

Setting	Suggested value	Why
assemble_mode	final_only	Assemble once when the loop is complete; much faster than recompositing every step.
assemble_device	auto or gpu	Uses GPU paste/weight accumulation when available.
base_canvas_mode	black when you do not want source pixels underneath	Prevents the original image from showing through uncovered or low-weight areas.
auto_composite_policy	safe_auto	Keeps background/context low and promotes focus/detail tiles.
small_tile_on_top	true for detail workflows	Helps small face/eye/text tiles win overlaps against larger body/context tiles.
mask_blend_mode	auto or mask_feather	Uses object masks when present and feathered rectangular masks otherwise.
color_correction	off first, then reinhard_lab, mkl_lab, or histogram as needed	Uses ComfyUI's official Transfer Color logic; reference defaults to the source image.

Detail tiles should blend over lower layers instead of cutting holes into them. The safe auto policy treats face, eyes, mouth, glasses, text, and similar focus regions as soft overlays when possible.

Recommended Starting Parameters

Scenario	Recommended settings
General manual grid	grid=3x3, default_pad=32, default_blend=32 or 64, round_to=8, include_full_image=false
Portrait/detail workflow	Auto Tile + Qwen/SAM prompt for person, face, eyes, hands, text, foreground object
Variable-size detail tiles	Use TTP Smart Tile Image Upscale Prep before sampler, then Output Size Estimate before Assemble
Avoid original image leaking through	base_canvas_mode=black
Small details should stay visible	small_tile_on_top=true, auto_composite_policy=safe_auto
Expensive alignment workflow	assemble_mode=final_only, connect done from Loop Collect
Strict manual layout	Keep Auto Tile mode as none after you finish editing, so formal runs do not replace your layout

Troubleshooting

Symptom	What to check
Auto Tile did not change the layout	Make sure auto_detect_mode is sam3.1 or qwenvl3, the required model input is connected, then read the editor status message.
QwenVL3 does not split the image	Confirm TTP QwenVL3 Local Loader is connected to qwen_vl_model. Qwen must return bbox JSON; Smart Tile accepts common bbox, bbox_2d, box_2d, objects, and xywh formats.
Running the full workflow replaces manual edits	Set auto_detect_mode=none after Auto Tile if you only wanted detection once, then manually edit the saved layout.
A manually moved mask tile still uses the old mask	Click Refresh masks after moving Grid-in child tiles from a masked parent.
Native Save Image saves multiple loop frames	Use TTP Smart Tile Save Final Image; loop previews are not treated as final output.
Face or eyes are hidden under a larger tile	Use TTP Smart Tile Semantic Rank, enable small_tile_on_top, and keep auto_composite_policy=safe_auto.
The original image shows through	Use base_canvas_mode=black and make sure Fill gaps has covered empty regions.
Assemble is slow	Use assemble_mode=final_only; avoid repeated pixel alignment during unfinished loop steps.

Advanced Layout JSON

Example layout:

{
  "defaults": {
    "pad": 128,
    "blend": 48,
    "priority": 50,
    "importance": 1.0
  },
  "tiles": [
    {
      "name": "full_image",
      "x": 0,
      "y": 0,
      "w": 1.0,
      "h": 1.0,
      "pad": 0,
      "blend": 96,
      "priority": 10,
      "importance": 0.5
    },
    {
      "name": "face",
      "x": 0.35,
      "y": 0.08,
      "w": 0.30,
      "h": 0.28,
      "pad": 192,
      "blend": 64,
      "priority": 100,
      "importance": 1.0
    }
  ]
}

Coordinates can be pixel values or normalized values from 0.0 to 1.0. A rectangle whose coordinates are all in 0..1 is treated as normalized, including browser-serialized 0 and 1 edges. pad is seam overlap: it expands only the tile edges that touch another tile. Outer canvas edges and non-adjacent gap edges are not expanded. blend, priority, and importance control how the sampled tile is pasted back.

For standard grid layouts, edge tiles are expanded inward with real source pixels when needed so the ComfyUI IMAGE batch has a consistent size without fake outer padding. Irregular manual layouts with uncovered gaps may still need transport padding because a single IMAGE batch cannot contain mixed image sizes.

1. Image Tile Batch Node

This node cuts an image into pieces automatically based on your specified width and height. It also records the necessary information for further processing.

Parameter	Description
Width	The width of each tile.
Height	The height of each tile.
Image	The image to be tiled.

Node View:

---

2. Image Assembly Node

This node reassembles image tiles back into a complete image while preventing visible lines between the tiles. It operates in pixel mode.

Parameter	Description
Tiles	Input the tiled image batch. Replace individual tiles if needed.
Position	Paired with the Image Tile Batch Node.
Original Size	Paired with the Image Tile Batch Node.
Grid Size	Paired with the Image Tile Batch Node.
Padding	The padding value used to merge the image pieces.

Node View:

---

3. Tile Image Size Node

This node calculates the resolution of each tile based on the original image dimensions and your specified width/height factors.

Parameter	Description
Width Factor	Divides the image width into equal parts.
Height Factor	Divides the image height into equal parts.

For example: A width factor of 2 and a height factor of 3 will divide the image into 6 equal tiles.

Node View:

---

4. Coordinate Splitter Node

This node converts position information into coordinates and connects them to the corresponding positions.

Node View:

---

5. Cond to Batch Node

This node converts condition lists into batches. It is reserved for future functionality expansion and connects to the conditions.

Node View:

---

6. Condition Merge Node

This node merges all tiled conditions into one and prepares them for building the final image. It connects to the Coordinate Splitter Node and Cond to Batch Node.

Node View:

---

Examples

Pixel Example (Recommended)

Latent Example

---

ControlNet Tile Integration

This workflow supports ControlNet Tile for enhanced upscaling. Here's an example of using tiles with the Hunyuan DIT model:

Resource	Link
Tile Example	Hugging Face Tile
Hunyuan 1.2	Download Hunyuan 1.2

Workflow Example:

---

Star History