ComfyUI Node
TBG ETUR Refiner PRO
Sampler and Refiner for TBG ETUR
TBG ETUR Refiner PRO
- model
- clip
- vae
- TBG_Pipe
- Controlnet_Pipe
- Enrichment_Pipe
- RF_UntwistingRoPE
- denoise_mask
- Redux_Style_Model
- Redux_Clip_Vision
- labs_refiner
- Refined
- Refined without Segs
- Refined without ColorCorrection
- Original Upscaled
- Original
- Tiles
◄model_typeFLUX1►
◄cfg1.0►
◄steps30►
◄seed4►
◄Flux_Guidance3.5►
◄sampler_name▾►
◄basic_scheduler▾►
◄vae_encodetiled slow►
◄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►
◄denoise_methodnormalized advanced►
◄Per_Pixel_Denoise_Mask_Strength1.00►
◄Image Stabilizer0.00►
◄Smoother - Sharper0.00►
◄Detail_Enhancer0.00►
◄Redux_strength0.000►
◄Controlnet_Pipe_strength1.00►
◄Color_MatchTBG ETUR Detail-Preserving Tile Stabilizer►
◄Scale-Invariant Feature Transformtrue►
◄Fast_1_Tile_Previewfalse►
◄Selected_Tiles_Onlyfalse►
◄Selected_Tiles_By_Numbers►
◄VRAM_ProfileLow VRAM Cache (Unload Models)►
CategoryTBG/ETUR Tiled Upscaler and Refiner
Inputs (36)
| 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 | COMBO | tiled slow | 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 | — |
| denoise_method | COMBO | normalized advanced | Default: splits sigmas by step percentage and interpolates back to full steps | Normalize Advanced: splits by sigma noise values and interpolates back to full steps, best for sampler-independent creative control | Default Short: splits sigmas by step percentage and keeps only low-noise steps for efficient img2img denoising. |
| Per_Pixel_Denoise_Mask_Strength | FLOAT | 1.000–1 | Changes the influence of the Per_Pixel_Denoise_Mask |
| Image Stabilizer | FLOAT | 0.000–1 | 0=OFF. Applys adaptive denoising based on local image complexity. Flat regions are stabilized to prevent color shifts, while detailed areas allow stronger creative changes. Useful for light backgrounds and large uniform areas and as alternative to cnets. |
| Smoother - Sharper | FLOAT | 0.00-1–1 | 0=OFF. Dual-stage adaptive sharpening. At high sigma (early steps), adds structured noise for detail invention. At low sigma (late steps), applies high-pass edge sharpening. Positive values sharpen and add details. Negative values soften and blur. Zero disables sharpening. Higher absolute values create stronger effects. |
| Detail_Enhancer | FLOAT | 0.00-1–1 | 0=OFF. Substep evaluation for detail control. Positive values (0.1-1.0): lookahead to next sigma, adds coherent details and refinement, reduces variation. Negative values (-0.1 to -1.0): lookback to previous sigma, adds creative variation and texture complexity. Zero = disabled (single pass, fastest). Performance cost: 2x slower on affected steps. |
| Redux_strength | FLOAT | 0.0000–1 | 0=OFF. It's a Redux multiplier value applied uniformly to all Tiles |
| Controlnet_Pipe_strength | FLOAT | 1.000–1 | 0=OFF. It's a multiplier value applied uniformly to all ControlNets from CnetPipe, scaling their combined influence. |
| 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. |
| 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 | |
| VRAM_Profile | COMBO | Low VRAM Cache (Unload Models) | Fast Cache (Max Speed): Precomputes full tile conditioning (text + Redux + ControlNet) for all tiles and keeps models loaded. Fastest sampling, highest RAM/VRAM usage. Low VRAM Cache (Unload Models): Same full precompute, then unloads models to reduce VRAM. RAM can still be high with many tiles. Ultra Low Memory (Per-Tile Streaming): Caches repeated text conditioning only; Redux/ControlNet are rebuilt per tile and released immediately. Also unloads/reloads models between steps/tiles for minimum VRAM. Slowest mode; best for very low-spec systems. |
| Controlnet_Pipeopt | Controlnet_Pipe | — | |
| Enrichment_Pipeopt | Enrichment_Pipe | — | |
| RF_UntwistingRoPEopt | RF_UntwistingRoPE_Pipe | — | |
| denoise_maskopt | MASK | — | |
| Redux_Style_Modelopt | STYLE_MODEL | — | |
| Redux_Clip_Visionopt | CLIP_VISION | — | |
| labs_refineropt | labs_refiner | TBG ETUR experimental tools |
Outputs (6)
| Name | Type | Description |
|---|---|---|
| Refined | IMAGE | — |
| Refined without Segs | IMAGE | — |
| Refined without ColorCorrection | IMAGE | — |
| Original Upscaled | IMAGE | — |
| Original | IMAGE | — |
| Tiles | IMAGE | — |