FLUX Studio

0
Model
Image Size
Seed
Result
Generated image
⬇ Download
◈ Image-to-Image: Upload a reference image, then describe what to refine or change. Lower denoise = closer to original.
🖼

Drop image here or browse

PNG, JPG, WEBP

Preview
Model
0
Denoise
0 = keep original · 1 = ignore original
0.60
Result
Refined image
⬇ Download

📖 Models Guide

Reference for RTX 3060 (12 GB VRAM) · ComfyUI + FLUX Studio

🖥 Your Setup — RTX 3060 (12 GB VRAM)

Model Size Fits? Notes
SD 1.5 ~2 GB ✓ Easy Legacy, fast, low quality. Good for testing workflows.
SDXL 1.0 ~7 GB ✓ Yes Good quality, needs 2-step (base+refiner). Popular LoRA ecosystem.
FLUX.1-schnell GGUF Q5 ~8 GB ✓ Optimal Current setup. Best quality/speed on 12 GB. Fast inference (~15s).
FLUX.1-dev (full) ~24 GB ✗ Too large Needs 24 GB+ VRAM. Use dev-GGUF Q3/Q4 (~10 GB) if needed.
Pony Diffusion XL ~7 GB ✓ Yes SDXL-based. Needs "score_9," style tags. Good for stylized art.

🔍 Where to Find Models

Hugging Face — huggingface.co
Best for FLUX models (black-forest-labs), CLIP/T5 encoders, VAEs. Direct download via web or huggingface-cli download. Free account needed for some gated models.
CivitAI — civitai.com
Best for SD 1.5, SDXL, Pony checkpoints and LoRAs. Filter by base model. Download via web or civitai-cli. Watch NSFW filter — toggle in account settings.
ComfyUI Manager
In ComfyUI web UI → Manager → Model Manager. Browse and install models directly. Also installs custom nodes.

📁 File Locations on This Machine

/mnt/raid0/comfyui/models/
├── unet/ ← FLUX GGUF checkpoints (flux1-schnell-Q5_K_S.gguf)
├── clip/ ← Text encoders (clip_l.safetensors, t5xxl_fp8_e4m3fn.safetensors)
├── vae/ ← VAE decoder (flux_ae.safetensors)
├── checkpoints/ ← SD 1.5 / SDXL / Pony full checkpoints (.safetensors)
├── loras/ ← LoRA adapters (match base model!)
├── controlnet/ ← ControlNet models
└── embeddings/ ← Textual Inversion files

⚠ FLUX GGUF checkpoints go in unet/, NOT checkpoints/. The workflow uses a different loader node (UnetLoader, not CheckpointLoader).

🔗 ComfyUI Workflow Node Structure

FLUX.1-schnell GGUF Workflow
UnetLoaderGGUF (unet/flux1-schnell-Q5_K_S.gguf)
DualCLIPLoader (clip_l + t5xxl_fp8_e4m3fn)
VAELoader (flux_ae.safetensors)
→ CLIPTextEncode × 2 (positive + negative)
→ EmptyLatentImage (width × height)
→ KSampler (steps: 4–6, cfg: 1.0, sampler: euler, scheduler: simple)
→ VAEDecode
→ SaveImage
SD 1.5 / SDXL Workflow
CheckpointLoaderSimple (checkpoints/model.safetensors)
→ CLIPTextEncode × 2 (positive + negative)
→ EmptyLatentImage
→ KSampler (steps: 20–30, cfg: 7.0, sampler: dpm++2m, scheduler: karras)
→ VAEDecode
→ SaveImage

💡 FLUX uses cfg=1.0 and very few steps (4–6). Standard SD settings (cfg 7, steps 25) will produce bad results with FLUX.

⚙ How to Install a New Model

  1. Identify model type and base: FLUX GGUF → unet/; checkpoint → checkpoints/; LoRA → loras/
  2. Download the file (from HuggingFace or CivitAI) to the correct folder on this machine under /mnt/raid0/comfyui/models/
  3. In ComfyUI web UI → click Refresh button (top-right) to reload model list — no restart needed
  4. Add the correct loader node for the model type and select the file from the dropdown
  5. LoRAs: add a LoraLoader node between the checkpoint loader and sampler; strength 0.5–1.0
Quick download example (SSH into .170):
cd /mnt/raid0/comfyui/models/checkpoints
wget -O sdxl_base.safetensors "https://..."
Or via huggingface-cli: huggingface-cli download stabilityai/stable-diffusion-xl-base-1.0 sd_xl_base_1.0.safetensors --local-dir ./

✍ Prompt Tips by Model

Model Prompt style Negative prompt
FLUX.1-schnell Natural sentences. No comma-spam. Describe scene fully. Not used (cfg=1.0)
SD 1.5 Comma-separated tags. Quality boosters: masterpiece, best quality blurry, bad anatomy, watermark
SDXL Natural sentences or tags. More coherent than SD 1.5. Similar to SD 1.5
Pony XL Must start with: score_9, score_8_up, score_7_up, score_4, score_3, score_2, score_1

✓ / ✗ Compatibility with FLUX Studio

Model / Family Folder Works here? Notes
FLUX.1-schnell GGUF unet/ ✓ Yes Current setup. Best for RTX 3060.
FLUX.1-dev GGUF (Q4/Q5) unet/ ✓ Yes ~10-12 GB, slower, higher quality than schnell.
SD 1.5 checkpoints checkpoints/ ✓ Yes DreamShaper, Realistic Vision, etc.
SDXL checkpoints checkpoints/ ✓ Yes Juggernaut XL, RealVisXL, DreamShaper XL, etc.
Pony Diffusion XL checkpoints/ ✓ Yes SDXL-based. Use score_9 tags in prompt.
SDXL-Turbo / SDXL-Lightning checkpoints/ ~ Partial Works but auto-detected steps/cfg may be wrong. Turbo: cfg=1, steps=1-4.
Kolors diffusion_models/ ✗ No Needs ChatGLM3 encoder + KolorsWrapper node. Use ComfyUI tab.
SD 3 / SD 3.5 checkpoints/ ✗ No Needs TripleCLIPLoader. Use ComfyUI tab.
HunyuanDiT / PixArt various ✗ No Custom loaders required. Use ComfyUI tab.
Video models (Wan, Hunyuan Video) various ✗ No Video generation — completely different pipeline.

⚠ Models marked ✗ appear in the dropdown with a warning label but generation will fail — use the ⬡ ComfyUI tab to run custom workflows for them.

💡 Best additions for RTX 3060: Juggernaut XL (SDXL, 7GB), DreamShaper XL (SDXL, 7GB), FLUX.1-dev GGUF Q4 (~10GB).