Tem um workflow do ComfyUI? Gere uma documentação legível direto no navegador.Torne um workflow legível.Experimentar agora ↗
Sexiam’s IMG2IMG + Upscale Workflow Guide (ComfyUI)
Perfil de execução
Descrição da fonte
Sexiam’s IMG2IMG + Upscale Workflow Guide (ComfyUI)
A clean, practical walkthrough for using the workflow effectively.
🧱 1. Load Your Checkpoint
Start here:
Load your SDXL checkpoint
(Optional) Load an external VAE
SDXL behaves differently depending on its VAE, so pick whichever looks best for your model.
🔧 2. Set Your VAE Mode
If you want to use the checkpoint’s built-in VAE, set:
Comentário gerado por IA
Explicação gerada por IA com base nos detalhes da fonte e da configuração. As sugestões são claramente identificadas.
Sexiam Img2Img 2.0 is a ComfyUI workflow for transforming a source image with SDXL 1.0 into a generated image-to-image result, with optional final model upscaling. Its inputs are a source image, positive and negative prompts, an SDXL checkpoint, a VAE choice, image orientation,
It loads the checkpoint and optional external VAE, loads the source image, checks portrait or landscape orientation, and uses a selected upscaler to resize the latent before sampling.
Before sampling, “Upscale by Model” resizes the latent for SDXL rather than enlarging the final image.
Positive and negative prompts control the generated image’s details, style, and adherence to the source image.
When denoise is 0.4 or lower, optional final model upscaling can enlarge the output by 1.5× or double its resolution at 2×.
Use it in ComfyUI with an SDXL 1.0 checkpoint; an external VAE is optional. Install the listed custom nodes through ComfyUI-Manager or manually via GitHub.
Listed model files are 4x_foolhardy_Remacri.pth, Nickel Saffron Manga.safetensors, and sdxl_vae.safetensors.
Listed custom node packs are ComfyLiterals, comfyui_essentials, comfyui-easy-use, comfyui-mixlab-nodes, masquerade-nodes-comfyui, and rgthree-comfy.
Set “Use Checkpoint VAE?” to True for the checkpoint’s built-in VAE, or False for an external VAE.
Set “Image is Portrait?” to True for portrait images and False for landscape images. The listed baselines are 832×1216 for portrait and 1216×832 for landscape, multiplied by 1.5×.
Use denoise 0.5 to preserve more composition and color, or 0.7 for a looser interpretation; lower values favor accuracy and higher values favor creativity.
Sugestão · não verificado
The input is used as a loose generational base, so some proportions and details may change.
Final upscaling results vary with GPU VRAM, system RAM, and upscale size.
Isso precisa ser editado?
Faça login para enviar uma solicitação de edição.
Fontes
1 fonteTrechos de fonte
1 trechoSource context: 345 downloads · Type Workflows · Base model SDXL 1.0
“Use Checkpoint VAE?” = True
If you want to use your external VAE, set:
“Use Checkpoint VAE?” = False
🖼️ 3. Load Your Input Image
Upload your source image using the Load Image node.
Orientation Check
Set the “Image is Portrait?” toggle correctly:
Portrait → True
Landscape → False
Load an Upscale Model
Pick the upscaler you want (Remacri, SwinIR, ESRGAN, etc.). You’re not upscaling the final output here — you’re shaping a clean latent size for SDXL.
📏 Why the Workflow Uses ‘Upscale by Model’ Before Sampling
You might ask: “Why am I using ‘Upscale by Model’ for image-to-image?”
Here’s the short answer:
➡️ You’re not actually upscaling to output — you’re resizing the latent. SDXL was trained on ~ 1MP images , so pushing too high causes:
Generation errors
Warped structure
Model collapse
This workflow uses a safe baseline:
832×1216 (portrait)
1216×832 (landscape)
Multiplied by 1.5× (50% larger)
This is the maximum size SDXL samplers can reliably handle before breaking.
If you're having issues, reset the “Scale to Target Ratio” nodes to whatever size works for your system.
This workflow is ideal for using the input as a loose generational base , meaning:
Great for creative reinterpretations
Fine for refinement
Some proportions/details may change (normal for IMG2IMG)
✍️ 4. Enter Your Prompts
Fill in your:
Positive prompt
Negative prompt
These control the details, style, and adherence to the original image.
🎛️ 5. Adjust KSampler Settings (Critical for IMG2IMG)
For IMG2IMG, denoise strength is the most important setting:
0.7 → Loose interpretation of the input
Model only reuses ~30% of the original
0.5 → Preserves composition + color
Allows detail refinement
Great for keeping the structure mostly intact
Use lower values when you want accuracy, higher when you want creativity.
🔍 6. Optional: Final Model Upscaling (Use Only at Low Denoise)
This section lets you upscale the final output if your denoise is 0.4 or lower .
Why this matters: Low denoise keeps most of the latent structure, so upscaling is stable.
Results vary based on:
GPU VRAM
System RAM
How large you upscale
Steps:
Load the final image
Set the upscale multiplier
1.5× → 50% bigger
2× → double the resolution
Choose an upscale model
Realistic outputs → clean general ESRGAN models
Stylized/anime outputs → anime-optimized upscale models
These are the only node packs used in the workflow. Install them through ComfyUI-Manager or manually via GitHub.
rgthree-comfy
comfyui-mixlab-nodes
Masquerade Nodes
ComfyLiterals
Estimativa de requisito VRAM
Estimativa indisponível
63,9 MB distribuídos em 1 de 3 arquivos de modelo. Total de arquivos do modelo + 25% de overhead de carregamento + 2 GB de buffer de execução, arredondado para cima.
Requisitos
Requisitos 9Upscaler · 63.9 MB · PTH · Unknown
Nickel Saffron Manga.safetensors
Não resolvidoCheckpoint · Unknown
sdxl_vae.safetensors
Não resolvidoVAE · Unknown
Pacote de nós · Registry
Pacote de nós · Registry
Pacote de nós · Registry
comfyui-mixlab-nodes
PossívelPacote de nós · Registry
Pacote de nós · Registry
Pacote de nós · Registry
