| name | flux-txt2img |
| description | Build Flux txt2img workflows with Flux.1 Dev (SRPO), Flux 2 Klein 9B, Turbo LoRAs, FluxGuidance, and DualCLIPLoader patterns |
| globs | ["**/*.json"] |
Flux Text-to-Image Workflows
Overview
Flux is a guidance-distilled diffusion model family from Black Forest Labs. It uses a separate FluxGuidance node instead of KSampler CFG (which must always be 1.0). Three variants are available locally:
- Flux.1 Dev SRPO. Fine-tuned Flux.1 Dev with SRPO alignment. Uses DualCLIPLoader (T5XXL + CLIP-L). BF16 only.
- Flux 2 Klein 9B. Distilled Flux 2 variant. Uses single CLIPLoader (Qwen3-8B) +
flux2-vae.safetensors. Fast 4-step generation.
- Flux 2 Turbo LoRA. Applied to Flux.1 Dev for 4-step generation.
Models
Flux.1 Dev SRPO
| Component | Node | Model | Notes |
|---|
| UNET | UNETLoader | flux.1-dev-SRPO-BFL-bf16.safetensors | 22.7GB, BF16 only — FP8 produces broken results |
| CLIP | DualCLIPLoader (type=flux) | clip_name1: t5xxl_fp8_e4m3fn.safetensors, clip_name2: clip_l.safetensors | T5XXL (4.7GB) + CLIP-L (235MB) |
| VAE | VAELoader | ae.safetensors | Standard Flux VAE (320MB). Z-Image uses the same VAE architecture but different weights — its VAE is a separate file (z-image-ae.safetensors), not this one |
Flux 2 Klein 9B
| Component | Node | Model | Notes |
|---|
| UNET | UNETLoader | bigLove_klein1.safetensors | 17.3GB, Klein 9B variant |
| CLIP | CLIPLoader (type=flux2) | qwen_3_8b_fp8mixed.safetensors | Qwen3-8B in text_encoders/ (8.3GB). Use flux2, NOT flux — both exist in the enum and flux fails at the sampler |
| VAE | VAELoader | flux2-vae.safetensors | Flux 2 specific VAE (321MB) |
Klein 9B vs Flux.1 Dev: Klein uses the Qwen3-8B text encoder (not T5XXL + CLIP-L). It has a different VAE (flux2-vae.safetensors). 9B distilled runs in 4 steps; 9B base needs ~50 steps at CFG 5.0. Fits in ~20GB VRAM with FP8.
Flux 2 Turbo LoRA (applied to Flux.1 Dev)
| Component | Node | Model | Notes |
|---|
| LoRA | LoraLoaderModelOnly | flux2-turbo-lora.safetensors | 2.6GB, strength 1.0 |
| Alt LoRA | LoraLoaderModelOnly | Flux2TurboComfyv2.safetensors | Community variant, same size |
Conditioning
CLIPTextEncodeFlux (Recommended for Flux.1 Dev)
Provides separate prompt fields for each text encoder:
{
"class_type": "CLIPTextEncodeFlux",
"inputs": {
"clip": ["<dual_clip>", 0],
"clip_l": "short prompt for CLIP-L",
"t5xxl": "detailed description for T5XXL",
"guidance": 3.5
}
}
clip_l captures key semantic features. t5xxl expands and refines descriptions. For simple use, put the same prompt in both fields. Guidance is built into this node, so no separate FluxGuidance is needed.
FluxGuidance (Alternative)
If using standard CLIPTextEncode instead of CLIPTextEncodeFlux, apply guidance separately:
{
"class_type": "FluxGuidance",
"inputs": {
"conditioning": ["<clip_text_encode>", 0],
"guidance": 3.5
}
}
Guidance Values
| Scenario | Guidance | Notes |
|---|
| Short prompts | 3.5–4.0 | Tighter prompt adherence |
| Long/complex prompts | 1.0–1.5 | More creative freedom |
| Realism | 2.5 | Less glossy skin, richer detail |
| Standard | 3.5 | Default for most use cases |
Negative Conditioning
Flux does not support traditional negative prompts (guidance-distilled, CFG=1.0). Use ConditioningZeroOut:
{
"class_type": "ConditioningZeroOut",
"inputs": { "conditioning": ["<positive_cond>", 0] }
}
Or use an empty CLIPTextEncode for the negative input.
Sampler Settings
Flux.1 Dev SRPO
| Parameter | Standard | Notes |
|---|
| steps | 20 | Range: 20–28 |
| cfg | 1.0 | Always 1.0 — guidance is via FluxGuidance |
| sampler_name | ipndm | Author-recommended for SRPO |
| scheduler | beta | Author-recommended for SRPO |
| guidance | 3.5 | Via CLIPTextEncodeFlux or FluxGuidance |
| denoise | 1.0 | |
The SRPO author recommends the ipndm/beta combo. Standard Flux settings (euler/simple) also work, but ipndm/beta gives better results with this fine-tune.
Flux 2 Klein 9B (Distilled)
| Parameter | Value | Notes |
|---|
| steps | 4 | Distilled model, 4 steps is optimal |
| cfg | 1.0 | Always 1.0 |
| sampler_name | euler | |
| scheduler | simple | |
| denoise | 1.0 | |
Flux 2 Klein 9B (Base/Undistilled)
| Parameter | Value | Notes |
|---|
| steps | 50 | Full quality |
| cfg | 5.0 | Higher CFG for base model |
| sampler_name | euler | |
| scheduler | simple | |
Flux.1 Dev + Turbo LoRA
| Parameter | Value | Notes |
|---|
| steps | 4 | Turbo-distilled |
| cfg | 1.0 | |
| sampler_name | euler | |
| scheduler | simple | |
| lora_strength | 1.0 | |
Resolutions
| Aspect | Resolution | Megapixels |
|---|
| Square | 1024x1024 | 1.0MP |
| Portrait 3:4 | 896x1152 | 1.0MP |
| Landscape 4:3 | 1152x896 | 1.0MP |
| Landscape 16:9 | 1344x768 | 1.0MP |
| Portrait 9:16 | 768x1344 | 1.0MP |
Flux operates at ~1 megapixel natively. Dimensions should be multiples of 8.
Prompt Style
Natural language descriptions. No quality tags needed (unlike SDXL/Illustrious). Detailed, descriptive prompts work best.
Good: "A young woman with auburn hair sits at a sunlit cafe in Paris, wearing a cream linen blazer, soft bokeh background, shot on Sony A7III 85mm f/1.4"
Bad: "masterpiece, best quality, 1girl, cafe, paris"
Complete Workflow: Flux.1 Dev SRPO
{
"1": { "class_type": "UNETLoader", "inputs": { "unet_name": "flux.1-dev-SRPO-BFL-bf16.safetensors", "weight_dtype": "default" }},
"2": { "class_type": "DualCLIPLoader", "inputs": { "clip_name1": "t5xxl_fp8_e4m3fn.safetensors", "clip_name2": "clip_l.safetensors", "type": "flux" }},
"3": { "class_type": "VAELoader", "inputs": { "vae_name": "ae.safetensors"
Complete Workflow: Flux 2 Klein 9B (Distilled, 4-Step)
{
"1": { "class_type": "UNETLoader", "inputs": { "unet_name": "bigLove_klein1.safetensors", "weight_dtype": "default" }},
"2": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen_3_8b_fp8mixed.safetensors", "type": "flux2" }},
"3": { "class_type": "VAELoader", "inputs": { "vae_name": "flux2-vae.safetensors" }},
"4"
Klein uses a single CLIPLoader (not DualCLIPLoader) with type: "flux2" and the Qwen3-8B text encoder from text_encoders/. The CLIP loader path resolves from models/text_encoders/.
Flux-2-specific gotchas, all of which fail at the KSampler rather than the loader, so the error points at the wrong node:
type must be flux2, not flux. Both values exist in the CLIPLoader enum, so flux loads without complaint and then dies during sampling.- Use
EmptyFlux2LatentImage, not EmptyLatentImage. Flux 2 uses a different latent channel count.
- Klein 9B pairs with the Qwen3-8B encoder (
qwen_3_8b* from Comfy-Org/vae-text-encorder-for-flux-klein-9b). The similarly-named qwen_3_4b ships in the klein-4b repo and is for the 4B model. Mismatching them raises mat1 and mat2 shapes cannot be multiplied (512x7680 and 12288x4096), where 7680 = 2560x3 (4B hidden size) and 12288 = 4096x3 (8B). It reads as a confusing CLIP error rather than a wrong-file error.
Complete Workflow: Flux.1 Dev + Turbo LoRA (4-Step)
{
"1": { "class_type": "UNETLoader", "inputs": { "unet_name": "flux.1-dev-SRPO-BFL-bf16.safetensors", "weight_dtype": "default" }},
"2": { "class_type": "LoraLoaderModelOnly", "inputs": { "model": ["1", 0], "lora_name": "flux2-turbo-lora.safetensors", "strength_model": 1.0 }},
"3": { "class_type": "DualCLIPLoader", "inputs":
LoRA Support
Custom LoRAs (jellyfish, etc.)
Apply Flux LoRAs with LoraLoaderModelOnly between UNET and KSampler:
{
"class_type": "LoraLoaderModelOnly",
"inputs": {
"model": ["<unet_or_previous_lora>", 0],
"lora_name": "<lora_file>.safetensors",
"strength_model": 1.0
}
}
Klein LoRAs
Klein 9B LoRAs go in the loras/Flux.2 Klein 9B/ subfolder:
klein_slider_detail.safetensors, a detail slider LoRA
VRAM Considerations
| Model | VRAM | Notes |
|---|
| SRPO BF16 + DualCLIP | ~24GB | Fills RTX 4090 exactly. Must use BF16 — FP8 is broken for SRPO |
| Klein 9B FP8 + Qwen3-8B | ~20GB | Fits comfortably on 4090 |
| SRPO + Turbo LoRA | ~24GB | Same as SRPO base |
- Always
clear_vram before switching to Flux from another model family
- T5XXL is the main VRAM consumer alongside the UNET; both stay loaded during sampling
- CLIP-L is small (235MB) and negligible
Tips
- KSampler CFG must always be 1.0. All guidance is through
CLIPTextEncodeFlux or FluxGuidance
- SRPO requires BF16. The FP8 quantization is known to produce broken results with this fine-tune
- For short prompts (1 to 2 sentences), increase guidance to 3.5 to 4.0. For long prompts (paragraph), decrease to 1.0 to 1.5
- Flux generates excellent text in images. Put text to render in quotes within your prompt
- Klein 9B is the fastest option at 4 steps. Use it for rapid iteration, then switch to SRPO for final quality
Sources
- Official: none found.
- Empirical: sampler values, wiring, and prompt notes from working graphs in
packs/ and observed renders; not a vendor prompting guide.
