ComfyUI Node
TBG ETUR Refiner CE
Sampler and Refiner for TBG ETUR
TBG ETUR Refiner CE
- model
- clip
- vae
- TBG_Pipe
- Controlnet_Pipe
- Enrichment_Pipe
- Redux_Style_Model
- Redux_Clip_Vision
- Refined
◄model_typeFLUX1►
◄cfg1.0►
◄steps30►
◄seed4►
◄Flux_Guidance3.5►
◄sampler_name▾►
◄basic_scheduler▾►
◄vae_encodetrue►
◄tile_size_vae1024►
◄General_Prompt_Positive►
◄General_Prompt_Negative低质量，模糊，噪点，失焦，曝光不良，过度曝光，欠曝光，重影，漂浮的物体，穿模，错误的结构，解剖错误，多余的肢体，多余的手指，缺少手指，手指融合，肢体融合，奇怪的骨骼，扭曲的身体，不自然的姿势，不自然的动作，不对称，身体比例不正确，脸部变形，重复的脸，五官错位，眼睛不对称，视线错误，面部畸形，表情僵硬，卡通化，非真实皮肤纹理，塑料感皮肤，过度光滑，噪点伪影，阴影错误，光照不一致，颜色溢出，奇怪的反射，重复的图案，破碎结构，AI 痕迹，水印，文字，logo，二维码，杂乱背景，物体穿插，图像缺损，像素化，低分辨率，乱色块，扭曲纹理，异常的毛发，不自然的布料褶皱，边缘锯齿，锐化过度，发光边缘，异常色彩，噪声纹理►
◄denoise0.27►
◄Color_MatchTBG ETUR Detail-Preserving Tile Stabilizer►
◄Scale-Invariant Feature Transformtrue►
◄Redux_strength0.000►
◄Fast_1_Tile_Previewfalse►
◄Selected_Tiles_Onlyfalse►
◄Selected_Tiles_By_Numbers►
CategoryTBG/ETUR Tiled Upscaler and Refiner
Inputs (26)
| Name | Type | Default | Description |
|---|---|---|---|
| model_type | COMBO | FLUX1 | 10 options: FLUX1, FLUX2, Ideogram4, FLUX1 Kontext, Qwen Image, Qwen Image Edit, +4 |
| model | MODEL | — | |
| clip | CLIP | — | |
| vae | VAE | — | |
| cfg | FLOAT | 1.0-10–100 | — |
| steps | INT | 301–10000 | — |
| TBG_Pipe | TBG_Pipe | — | |
| seed | INT | 40–18446744073709550000 | — |
| Flux_Guidance | FLOAT | 3.5-100–100 | All Fusion Modes benefit from high Guidance, so if you notice that certain areas aren't blending well, try increasing the Guidance value. |
| sampler_name | COMBO | 47 options: euler, TBG Flux2 Sampler, pid_sde, pid_creative_sde, euler_cfg_pp, euler_ancestral, +41 | |
| basic_scheduler | COMBO | 9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3 | |
| vae_encode | BOOLEAN | true | tiled slow is the standard ComfyUI tiled VAE path. tbg Color-preserving fast is the faster TBG VAE path that preserves colors. Nvidia PiD 4x changes VAE decoding to the PiD model; 1024x1024 uses the native fast path, while other tile or segment sizes use tiled PiD latent decode. It works with FLUX1, FLUX2, Qwen Image, Qwen Image Edit, SDXL, SD3, and Z-Image. |
| tile_size_vae | INT | 1024256–4096 | — |
| General_Prompt_Positive | STRING | General_Prompt_Positive | |
| General_Prompt_Negative | STRING | 低质量，模糊，噪点，失焦，曝光不良，过度曝光，欠曝光，重影，漂浮的物体，穿模，错误的结构，解剖错误，多余的肢体，多余的手指，缺少手指，手指融合，肢体融合，奇怪的骨骼，扭曲的身体，不自然的姿势，不自然的动作，不对称，身体比例不正确，脸部变形，重复的脸，五官错位，眼睛不对称，视线错误，面部畸形，表情僵硬，卡通化，非真实皮肤纹理，塑料感皮肤，过度光滑，噪点伪影，阴影错误，光照不一致，颜色溢出，奇怪的反射，重复的图案，破碎结构，AI 痕迹，水印，文字，logo，二维码，杂乱背景，物体穿插，图像缺损，像素化，低分辨率，乱色块，扭曲纹理，异常的毛发，不自然的布料褶皱，边缘锯齿，锐化过度，发光边缘，异常色彩，噪声纹理 | General_Prompt_Negative |
| denoise | FLOAT | 0.270–1 | — |
| Color_Match | COMBO | TBG ETUR Detail-Preserving Tile Stabilizer | Tile-aware detail-preserving color stabilization with feathered seam blending into generated neighbor tiles. If tile-aware conditions are unavailable, it falls back to TBG Detail-Preserving Color Stabilizer. |
| Scale-Invariant Feature Transform | BOOLEAN | true | After a tile is generated, aligns it back to the reference tile. Off disables it. On uses the strongest x4 drift-correction path. |
| Redux_strength | FLOAT | 0.0000–1 | — |
| Fast_1_Tile_Preview | BOOLEAN | false | The first Selected_Tiles_By_Number are processed at full scale as a preview, allowing a quick check of settings before processing the entire set. |
| Selected_Tiles_Only | BOOLEAN | false | — |
| Selected_Tiles_By_Numbers | STRING | You can set a list of selected tiles to process like 1,2,3,6 and activate Selected_Tiles_Only | |
| Controlnet_Pipeopt | Controlnet_Pipe | — | |
| Enrichment_Pipeopt | Enrichment_Pipe | — | |
| Redux_Style_Modelopt | STYLE_MODEL | — | |
| Redux_Clip_Visionopt | CLIP_VISION | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| Refined | IMAGE | — |