Generated image with BHS Lora, Upscaled Latent, not a real actor or character

This image made in the workflow Olemara's High Quality 2pass Qwen 2.1 Image Workflow v.1.01 and
based on generated before image below

High Quality 2pass-Upscaling Qwen 2.1 Image Workflow v.1.01

Why this workflow

Most people still bolt old face-swap stacks onto Qwen (InstantID-style helpers, separate swap pipelines, extra detectors). On Qwen-Image 2.1 that is mostly unnecessary.

This graph is a Turbo DiT setup with an experimental but stable LoRA stack, a two-pass latent upscale, and one clear job in the default prompt path:

*<image1> = full base frame (body, light, background)**

*<image2> = donor head**

BHS (Best Head Swap) LoRA does the head replacement inside the same generate — no second tool.

On top of that:

- Lightning / turbo LoRA — native few-step sampling (about 4–8 steps)

- Multiple Angles LoRA — <mva> + azimuth / elevation labels (camera orbit without reshooting the ref)

- Detail enhancer LoRA — enhance this image / 1_enhance this image

- Natural Exposure LoRA — more believable light and skin response

- Optional NSFW / MysticXXX LoRAs when you need adult anatomy without fighting the base model

- Wildcard prompt node so one graph can randomize angle + shot height every run

- Texture-fix VAE + 2-pass latent upscale so the second pass tightens skin, edges, and micro-contrast instead of only blowing up pixels

Free to use. If something breaks on your machine, ask in the comments — happy to help.

---

Core idea (head swap done right)

1. Load base as <image1> (pose, wardrobe, room, lighting).

2. Load face/head donor as <image2>. Optionaly - background <image3>

3. Keep the head_swap block in the prompt (see below).

4. Run. BHS LoRA strength usually 1.0; if it copies the whole ref instead of swapping, drop to ~0.8.

You are not running a separate “face replace” app. The diffusion pass is the swap.

---

Models (what the graph expects)

• DiT / UNET | qwen-image-2.1-turbo-int8_convrot.safetensors or …-fp8.safetensors - Turbo quant from Qwen-Image-2.1-Turbo-Quantized, Can be used with 8 steps without any lightning loras
  Text encoder | qwen3vl_8b_nvfp4_heretic.safetensors | CLIP type: qwen_image (Abliterated / Heretic NVFP4) |

• VAE | texture_fix_vae_for_qwen_image_2.1_bf16.safetensors | Prefer this over a random VAE |

LoRAs used in the workflow

• Lightning (few-step) | pytorch_lora_weights.safetensors | Optional Enables stable 4–8 step turbo sampling |

• BHS – Best Head Swap | bfs_head_v1.1_alternative_qwen_2.1.safetensors | head_swap: block; strength 1.0 → 0.8 if needed |

• Detail enhancer | elusarcas-qwen2-1-detailer-v1.safetensors | enhance this image / 1_enhance this image |

• Multiple angles | multiple_angles_step1000.safetensors | Start prompt with <mva> + labeled angle |

• Natural Exposure | Qwen-Image-2.1-Natural-Exposure-LoRA-4000.safetensors | Natural light / exposure response |

• MysticXXX (optional) | Qwen-MysticXXX-v1.safetensors | Adult / denser detail when needed |

• NSFW 2.1 (optional) | nsfw_lora_2.1_v2.safetensors | Adult anatomy helper |

---

Default wildcard-style prompt (as in the graph)

Uses [custom-willdcard-prompt](https://github.com/eliseipanov/custom-willdcard-prompt) so {a\|b\|c} picks one option per run:

<mva> {front view 0°|front-right view 30°|front-right quarter view 60°|right side view 90°|back-right quarter view 120°|back-right view 150°|back view 180°|back-left view 210°|back-left quarter view 240°|left side view 270°|front-left quarter view 300°|front-left view 330°}, {high-angle shot|top-down shot|eye-level shot}.

head_swap: start with <image1> as the base image, keeping its lighting, environment, and background. Remove completely the head of the Woman from <image1> and replace it with the head of the Woman from <image2>, strictly preserving the hair, eye color, nose structure from <image2>. Copy the direction of the eye, head rotation, micro expressions from <image1>, high quality, sharp details, 4k. Keep the colour scheme of <image1>

camera and lighting: Cinematic realistic lighting, soft even light, realistic skin texture, realistic skin highlights, high detail. No text, no logos, no watermark. Shot made by fullframe Nikon camera with prime lens 85mm f/1.8. film grain, Dramatic lighting, cold long shadows.

enhance this image: make all details of skin better, improve micro contrast, the point of black and point of white, sharp manual focus and white balance

*<mva> rules (Multiple Angles LoRA)**

- Put <mva> at the start of the prompt.

- Use labels only (not raw degrees like 47°).

- Azimuth examples: front view 0°, right side view 90°, back view 180°, …

- Elevation: eye-level shot, elevated shot 30°, high-angle shot, top-down shot.

- Optional: add close-up at the end of the angle line.

- LoRA strength ~**0.8–1.0** (lower if it pastes the ref instead of changing camera).

BHS tip: if the head “sticks” wrong, lower BHS toward 0.8 and keep the head_swap: sentence explicit about hair / eyes / nose from <image2> and pose / gaze from <image1>.

---

Sampler / quality (practical)

- Turbo model in workflow can be used as is with 8 steps.
- Turbo model + Lightning Lora: about 4-5 steps, low CFG (graph example ~**1.1**), euler + beta is a known-good combo on this stack. You also can use er_sde + sgm_uniform or Euler_ancestral / Beta *(or sgm_uniform)

- Always run the second pass / latent upscale if you care about skin pores, wet highlights, fabric, hair edges

- Output size: keep multiples of 32; ResolutionSelector-style ~1.5 MP 3:2 is a solid default before upscale

---

Nodes you need

- ComfyUI build that includes TextEncodeQwenImage21 (Qwen-Image 2.1 path)

- UNETLoader / CLIPLoader qwen_image) / VAELoader

- LoadImage ×2 minimum <image1> base, <image2> head)

- KSampler (+ second pass / latent upscaler node and model as in the graph)

- Wildcard / custom-willdcard-prompt (or paste a fixed angle line)

- Optional: LUT-from-reference (Just load the image you want you generated photo looks like by style) + light grade nodes if you keep the color group

---

Honest limitations

- Quantized Turbo is fast, not magic — garbage refs in → garbage swap out

- <mva> wants named angles; random degree numbers are unreliable

- Stacking every LoRA at 1.0 can muddy the image; if detail collapses, turn optional NSFW/Mystic down first, not BHS

- VRAM: INT8/FP8 Turbo is the point; the 2nd latent pass + texture VAE is what usually hurts small GPUs

---

License / tone

Workflow and notes are shared free for the community. Remix it, post results, tag the LoRA authors. Questions welcome in the comments — especially if BHS or <mva> misbehave on your exact quant.

---

### Workflow Notes (short install list)

Weights

1. DiT: qwen-image-2.1-turbo-int8_convrot.safetensors or FP8 twin → models/diffusion_models

2. TE: qwen3vl_8b_nvfp4_heretic.safetensors → models/text_encoders (type qwen_image)

3. VAE: texture_fix_vae_for_qwen_image_2.1_bf16.safetensors → models/vae

4. LoRAs → models/loras/:

- pytorch_lora_weights.safetensors (Lightning Lora from )

- bfs_head_v1.1_alternative_qwen_2.1.safetensors (BHS)

- elusarcas-qwen2-1-detailer-v1.safetensors

- multiple_angles_step1000.safetensors

- Qwen-Image-2.1-Natural-Exposure-LoRA-4000.safetensors

- optional: Qwen-MysticXXX-v1.safetensors, nsfw_lora_2.1_v2.safetensors

Custom nodes

- Qwen-Image 2.1 encoder support (stock recent Comfy or pack that ships TextEncodeQwenImage21)

- [custom-willdcard-prompt](https://github.com/eliseipanov/custom-willdcard-prompt) for {a|b|c}

- Whatever upscale/2pass nodes are already wired in the uploaded graph

Run order

1. <image1> base, <image2> head, also optional (disabled) <image3> for background

2. Confirm BHS + Lightning + (optional) angles/detail/exposure strengths

3. Generate few-step turbo → let 2pass latent upscale finish

4. Only then push grade/LUT

Support:

comment with GPU, quant (INT8/FP8), and which LoRA strengths you used if the head swap or <mva> drifts.