Flux Kontext + PuLID + Detail Daemon — Character LoRA Dataset Generator (164+ images)

Running Kontext ReferenceLatent conditioning through standard PuLID throws an error about

unexpected conditioning format. Fix: patch pulidflux.py to add timestep_zero_index kwarg

support and generation-token-only injection mode. Full patch instructions in the linked Article.

━━━ HARDWARE REQUIREMENTS ━━━

• Minimum: 16GB VRAM (will be slow, may need to reduce steps/resolution)

• Recommended: 24GB+ VRAM (RTX 3090/4090)

• Ideal: 48–96GB VRAM (A100, H100, RTX 6000 Ada)

• RAM: 32GB+ system RAM

• Storage: ~22GB for all models + 5–10GB per batch run output

━━━ REQUIRED MODELS ━━━

Place in the specified ComfyUI subdirectory:

• models/unet/ — flux1-dev-kontext_fp8_scaled.safetensors (~17GB) [HuggingFace: black-forest-labs]

• models/vae/ — ae.safetensors (~335MB) [HuggingFace: black-forest-labs/FLUX.1-dev]

• models/clip/ — clip_l.safetensors (~246MB) + t5xxl_fp8_e4m3fn_scaled.safetensors (~4.7GB)

• models/loras/ — flux_realism_lora.safetensors (XLabs — CivitAI)

• models/loras/ — fluxRealSkin_v2.safetensors (CivitAI)

• models/pulid/ — pulid_flux_v0.9.0.safetensors

InsightFace weights download automatically on first run (internet required for first execution only).

━━━ REQUIRED CUSTOM NODES ━━━

• ComfyUI-PuLID-Flux (with pulidflux.py patch — see Article)

• ComfyUI-Detail-Daemon

• ComfyUI-Kontext (ReferenceLatent + FluxKontextImageScale — built into recent ComfyUI)

━━━ TO USE ━━━

1. Place 3 reference images in ComfyUI/input/:

— Node 4: full/3-quarter body reference

— Node 30: head+shoulders face reference

— Node 20: tight face crop for PuLID (face fills 70%+ of frame, near-frontal angle)

2. Edit CLIPTextEncode (Node 7) with your character description

3. Run a single image first to verify setup, then queue the full batch

━━━ KEY PARAMETERS ━━━

PuLID — Node 24:

weight 0.73 (0.35–0.75 useful range) — higher = stronger identity, lower = more variation

start_at 0.05 — leave at 0.05; lets structure form before identity locks

end_at 0.72 — lower if skin looks plastic; releases final texture passes from PuLID constraint

Detail Daemon — Node 50:

detail_amount 0.40 (0.20–0.55 range) — sweet spot; 0.50+ = over-sharpened

start 0.25 / end 0.80 — active across mid-sampling where detail resolves

LoRA strengths:

XLabs Realism (Node 40): 0.60 (0.35–0.75 useful range)

FluxRealSkin v2 (Node 41): 0.45 (0.25–0.55 useful range)

Note: these two compound — if you raise one, slightly lower the other

Sampling: 25 steps, euler_ancestral, simple scheduler, guidance 2.5

━━━ FILES INCLUDED ━━━

• workflow_ui_format.json — import directly into ComfyUI

Generates 164 photorealistic identity-consistent images from 3 reference photos.

Built for high-quality synthetic training datasets for Wan 2.2 / HunyuanVideo character

LoRAs — but outputs are also great standalone character content. Developed across 16

test iterations for maximum photorealism, identity consistency, and anatomical accuracy.

━━━ WHAT IT DOES ━━━

Combines three techniques that individually exist in the community but haven't been

documented together in a single production-ready batch pipeline:

• Flux Kontext DUAL reference chain — body ref + face ref simultaneously, stronger than single ref

• PuLID Flux face identity injection — biometric-level face locking via InsightFace embeddings

• Detail Daemon sampler augmentation — forces skin pores, hair strands, iris texture to resolve

• XLabs Flux Realism LoRA + FluxRealSkin v2 stacked for photographic skin rendering

━━━ OUTPUT ━━━

82 prompts × 2 seeds = 164 images:

• 10 extreme close-ups (frontal, 90° profile left/right, dutch tilt, high angle, low angle, head down)

• 15 portraits head+shoulders (laughing, looking away, over-shoulder, near-profile, editorial, candid)

• 15 medium waist-up shots

• 12 full body — 768×1152 PORTRAIT canvas so subjects look tall, not short

• 12 revealing/fashion outfits

• 12 lifestyle/candid

• 6 lighting moods (golden hour, dramatic shadow, neon night, overcast, backlit, morning window)

━━━ KEY TECHNICAL FINDINGS ━━━

1. PuLID must be applied AFTER LoRAs — LoRAs define rendering style first, PuLID locks face into it

2. Detail Daemon wraps the SAMPLER output of KSamplerSelect, NOT the sigmas — wrong wiring = 400 error

3. SamplerCustomAdvanced + BasicGuider is REQUIRED when using PuLID (standard KSampler won't work)

4. 768×1152 portrait canvas for full body — square 1024×1024 makes subjects look short/wide

━━━ COMPATIBILITY WARNING — READ BEFORE INSTALLING ━━━

The standard ComfyUI-PuLID-Flux node is NOT compatible with Flux Kontext out of the box.