✨ Qwen-Image 2.1 Glamour Portrait — 4-Mode Auto-Prompt Workflow (SFW)
ComfyUI · Qwen-Image 2.1 · one control panel · Built with Qwen
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Pick a MODE, press Queue, get a finished 4:5 portrait. The workflow writes its own prompt: a built-in list of scene ideas (200 per mode) is expanded by Qwen3-VL 8B running inside ComfyUI into a detailed photographic paragraph, a clothing guard keeps every image inside the MODE you chose, and Qwen-Image 2.1 renders it. No LoRA is included or needed. All subjects are fictional, AI-generated adults. Tested on an RTX 5080 16 GB.
🔀 Which version should I download?
• v1.2 = the standard one. Only the three model files below, 25 steps, about 70 s per image on an RTX 5080 16 GB. Nothing extra to install.
• v1.3 = Viggle Fast. Same panel, same prompts, but the sampler is replaced by Viggle Turbo (6 steps): about 22 s per image on the same card, about 3x faster. It needs two extra files (the Viggle model file and one small custom node, see below) and was tested on a 16 GB card only. If you do not want to install anything extra, keep using v1.2.
⚡ Speed at a glance (RTX 5080 16 GB, 1792×2240, MODE 1–4, same prompts and seeds on both)
• v1.2, 25 steps: about 70 s per image (median of 12 images: 69.7 s).
• v1.3 Viggle, 6 steps: about 22 s per image (median of 12 images: 21.9 s) — about 3x faster. The first image of a session is slower while the models load (50 s in our run).
• v1.3 has no FAST switch: it always renders in 6 steps. The FAST switch (v1.1 and v1.2) is described under Version History.
• Smaller images render faster in v1.2 too: set the Empty Latent Image size to 1280×1600 (about half the pixels of the default). For v1.3 we only tested 1792×2240.
💻 Requirements
v1.3 (Viggle Fast):
• GPU memory: 16 GB tested. Peak device memory in our runs was about 15 GB (14,966 MiB) on a 16 GB card. We did not test 12 GB or anything below 16 GB with v1.3, so do not assume it fits.
• System RAM: 32 GB (same as v1.2, not re-measured for v1.3).
• ComfyUI v0.37.0 or newer (we ran 0.37.1), plus the two extra files described under Required Models and Required Custom Nodes.
• Image size: tested at 1792×2240 only. The Viggle sigma node adjusts to the image size by itself, but we did not check image quality at other sizes.
• Upscale (SeedVR2) and Fix: not tested with v1.3.
v1.2 (standard): 12 GB+ (recommended 16 GB), system RAM 32 GB.
• What we measured for v1.2: on a test build of this workflow (it had the optional Fix LoRA loaded; the release file does not load it), ComfyUI was started with --reserve-vram 4 to EMULATE a 12 GB card. Peak device memory was 13.2–14.2 GB and an image took 121–180 s. That was an emulation on a 16 GB card, not a physical 12 GB card, so we do not claim it needs less.
• 8 GB: expect it to be slow; we have not tested it. Below 8 GB: not tested.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✨ Features
✅ One CONTROL PANEL, top-left — MODE, subject source, upscale switch, seed. Everything else is "Engine — no need to touch"
✅ 4 MODEs — 1 Everyday · 2 Fitted outfit · 3 Swimwear · 4 Lingerie (opaque, lined). Clothing guards are built in
✅ Random from the built-in list, or "My own idea" — type your own scene line; the expander and guards still apply
✅ v1.3: Viggle Fast — Viggle Turbo v0.3, 6 fixed steps, about 3x faster than 25 steps in our tests (two extra files needed)
✅ FAST switch (v1.1 and v1.2 only) — 0 Quality (25 steps, default, no LoRA) · 1 Fast (16 steps, no LoRA). Replaced by Viggle in v1.3
✅ BEAUTY switch — 0 Natural (default) · 1 Beauty (adds one line to the expander for a more polished, clearly adult look). Present in v1.3 as in v1.2; we ran the v1.3 tests with BEAUTY at 0
✅ Auto-prompt expander inside the graph — Qwen3-VL 8B (same file as the text encoder, no extra download)
✅ Hands-friendly prompt wording — simple hand anchors, no finger descriptions
✅ Optional Upscale — SeedVR2 2× (1792×2240 → 2560×3200), OFF by default (not tested with v1.3)
✅ Empty "Optional LoRA" slot — bypassed; drop in your own LoRA if you want one (see our Realism LoRA below)
✅ Civitai metadata embedded — model hash, prompt, seed, sampler, steps in every PNG (Image Saver). A LoRA you switch on yourself is NOT written into the metadata automatically
✅ No external files — all 4 subject lists are inside the workflow
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📦 Required Models
v1.2 (3 files, ≈17.3 GB):
• qwen_image_2.1_int8_convrot.safetensors (7.26 GB) — diffusion model → ComfyUI/models/diffusion_models/
• qwen3vl_8b_int8_convrot.safetensors (9.35 GB) — text encoder AND prompt expander → ComfyUI/models/text_encoders/
• qwen_image_2.1_vae_bf16.safetensors (0.68 GB) — VAE → ComfyUI/models/vae/
v1.3 (3 files, ≈17.3 GB): the same text encoder and VAE as above, plus the Viggle file instead of qwen_image_2.1_int8_convrot.safetensors:
• Qwen-Image-2.1-viggle-turbo-v0.3-6step-int8_convrot.safetensors (7.26 GB) — diffusion model with the 6-step Viggle LoRA already merged in → ComfyUI/models/diffusion_models/
Optional upscale (only if you turn Upscale ON, v1.2):
• seedvr2_ema_7b_fp16.safetensors (16.48 GB) + ema_vae_fp16.safetensors (0.50 GB) → ComfyUI/models/SEEDVR2/
• Lighter option: seedvr2_ema_3b_fp16.safetensors (6.78 GB) — in our 2-image check it matched the 7B on face and skin detail at 100% and ran at the same speed, a good choice if you have less RAM
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⬇️ Download Links (Hugging Face)
• https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/diffusion_models/qwen_image_2.1_int8_convrot.safetensors (v1.2)
• https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/text_encoders/qwen3vl_8b_int8_convrot.safetensors
• https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors
• Viggle model file (v1.3): https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/resolve/main/Qwen-Image-2.1-viggle-turbo-v0.3-6step-int8_convrot.safetensors
• Viggle repo page (licence, notice, readme): https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo
• SeedVR2 (optional): https://huggingface.co/numz/SeedVR2_comfyUI/tree/main
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🧩 Required Custom Nodes (items 1–6 are installable with ComfyUI-Manager; the Viggle node for v1.3 is NOT — install it by hand)
1. WAS Node Suite — https://github.com/WASasquatch/was-node-suite-comfyui
2. ComfyUI-Impact-Pack — https://github.com/ltdrdata/ComfyUI-Impact-Pack
3. rgthree-comfy — https://github.com/rgthree/rgthree-comfy
4. ComfyUI_essentials — https://github.com/cubiq/ComfyUI_essentials
5. ComfyUI-Image-Saver — https://github.com/alexopus/ComfyUI-Image-Saver
6. ComfyUI-SeedVR2_VideoUpscaler (only for the optional Upscale) — https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler
7. v1.3 only — viggle_turbo.py (by Viggle, 5.5 KB, adds the node "Viggle Turbo Sigmas"):
• Download it from https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/resolve/main/comfyui/viggle_turbo.py
• Copy it into ComfyUI/custom_nodes/ and restart ComfyUI. Until you do, v1.3 shows ViggleTurboSigmas as a missing node.
• It is third-party code. Its author says the ComfyUI port is largely written with an AI assistant and may have rough edges. We read the whole file (repo revision 009a44a8): it has no network, subprocess or file-writing calls, and its SHA-256 is 0592defd8ee6555b8c02bd9b287514849b1d46f579f47f8c7f711e6411f89587. Please read it yourself before you install it.
Everything else is ComfyUI core. ComfyUI v0.37.0 or newer is required (Qwen-Image 2.1 nodes).
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🚀 How to Use
1. Download the model files for your version into the folders above. The loaders already use these exact file names. For v1.3, also install viggle_turbo.py as described above and restart ComfyUI.
2. In the CONTROL PANEL pick MODE: 1 Everyday · 2 Fitted outfit · 3 Swimwear · 4 Lingerie.
3. Press Queue. A new random image is made every run.
Optional:
• Subject source = 2 and type a scene in "My own idea" (for example: full body, red satin dress, rooftop bar at golden hour)
• Upscale: tick it in the switch on the left
• Several images: raise Batch count next to the Queue button
• Person box: the default is "Korean woman in her mid 20s" — change it to any adult you like
• FAST (v1.1 and v1.2): 0 Quality (default) · 1 Fast (fewer steps, no download needed). Not in v1.3.
• BEAUTY: 0 Natural (default) · 1 Beauty — no extra download
• Optional LoRA (v1.3 example images use it): see the next block.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🎨 Optional: our Realism LoRA
Qwen-Image 2.1 Glamour Realism LoRA v1.0 — https://civitai.com/models/2986100 . Put the file in ComfyUI/models/loras, pick it in the "Optional LoRA" node and un-bypass that node. Recommended strength 0.6 (0.5–0.7 works; 0.8 is not recommended because outfits and poses drift). There is no trigger word.
• What we tested: v1.2 at 25 steps, and v1.3 (Viggle) with 4 images (one seed per MODE, strength 0.6) plus the 24-image batch behind the v1.3 example images — no anatomy problems found in the 4-image test. Only strength 0.6 was tested with v1.3.
• The workflow does not include the LoRA, and the node is empty and bypassed by default.
• If you switch it on, the PNG metadata does not list the LoRA by itself — add it as a resource when you post.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚙️ Settings & Parameters
• Size: 1792×2240 (4:5 portrait)
• v1.2: sampler euler / simple · steps 25 (FAST 16) · CFG 1.0
• v1.3: Viggle Turbo sigmas, sampler euler, 6 steps, no CFG, no negative prompt (the old negative text is not used)
• Seed: randomize each run (switch to "fixed" in the Seed node to repeat an image)
• Expander: Qwen3-VL 8B, temperature 0.7, max 768 tokens, plus 5 small clean-up steps (3 text clean-ups, a refusal / cut-off guard and a subject-clause trim)
• Output: PNG with Civitai metadata, saved to ComfyUI/output/qwen21_glamour/<mode>/
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
💡 Performance Tips
• v1.2 base image 1792×2240: about 70 s on an RTX 5080 16 GB (first run is slower while models load). v1.3: about 22 s. Upscale adds about 30 s (v1.2 measurement).
• Upscale is OFF by default so first runs fit the requirements above with 32 GB RAM. The 7B upscale model needs about 48 GB RAM; if you turn it on with less, use the 3B model.
• Pairs well with "Qwen Image 2.1 Fix v1.0" by _Envy_ (optional, not included, not needed, not tested with v1.3): https://civitai.com/models/2957332
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📋 Version History
• v1.0 — first release: 4-mode auto-prompt workflow, one control panel, no LoRA needed.
• v1.1 — adds two panel switches: FAST (0 Quality, 25 steps, no LoRA, about 68 s · 1 Fast, 16 steps, no LoRA, about 40 s) and BEAUTY (0 Natural · 1 Beauty, adds a polished-look line to the prompt, about 70 s, same cost as Quality). Resolution is 1792×2240 at every FAST level (1280×1600 is still available as a manual Empty Latent Image tip, see Performance Tips). Defaults unchanged (FAST=0, BEAUTY=0 render identically to v1.0). Tested on 16 cases x 2 seeds x 2 pages at 100% crop (32 Fast-arm renders, 16 Beauty-arm renders): 30/32 Fast renders (14/16 cases) were clean at 16 steps; 2/32 (both MODE1 test seeds) showed occasional ghosting/doubling on a limb or edge, an effect confirmed absent at Quality's 25 steps for the same seeds — use Quality if you need a guaranteed-clean result. Beauty read as more polished and clearly adult in all 16 cases.
• v1.2 — text-path update only; models, sampler, steps and the FAST / BEAUTY switches are unchanged. The built-in scene lists now describe neutral, open poses (no kneeling, folded-arm or head-back lines). One directional-light sentence was added to the prompt expander. A guard falls back to the raw scene line when the expander refuses or cuts off, and a clean-up step removes subject re-statements, skin-texture words and invented hand placements. The expander limit went from 400 to 768 tokens, skin wording in the prompts is positive-only (matte, dry, smooth, even), and hands are described only when the scene line does. The Optional LoRA slot is still empty and bypassed.
• v1.3 — Viggle Fast. Same prompts, panel and text path as v1.2; only the sampler changes. The 25-step KSampler is replaced by Viggle Turbo v0.3 (6 steps, euler, no CFG), so the FAST switch is gone. Needs the Viggle model file (replaces the v1.2 diffusion model) and the viggle_turbo.py custom node. Image Saver now writes Steps 6 and the Viggle model name and hash into the PNG. On the same 12 prompts and seeds at 1792×2240 it ran in about 22 s per image against about 70 s for v1.2 (about 3x), and in our own review all 12 images passed with no limb or anatomy faults, like the 12 v1.2 images. We also ran the release file itself (4 images, MODE 1–4, defaults): 4 of 4 passed our review. Small samples, one seed per case. Skin highlights can look a little glossier than at 25 steps in some images. Tested at 1792×2240 on a 16 GB card only; 12 GB, other image sizes, Upscale and Fix were not tested.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🖼️ About the example images
• Examples 1–15 come from v1.0–v1.2.
• Examples 16–27 were made with v1.3 plus our optional Realism LoRA at strength 0.6: 12 images picked by us from one batch of 24 (4 MODEs x 6 seeds, one generation each at 1792×2240, no upscaler, no face fix, no inpainting). One of the 24 was rejected by our own check (the face did not match the intended look); we saw no anatomy problems in the others. Because we picked the best 12, treat them as a best-case sample, not an average.
• Each example PNG carries the ComfyUI workflow it was made with. That embedded workflow is the LoRA-on variant, not the downloadable file (where the LoRA node is bypassed and empty): drag a PNG in, then you also need the LoRA file and the Viggle files.
• The PNG metadata lists the LoRA. If you post your own images made with the LoRA, add it as a resource yourself.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔗 Cross-links
Adult version (18+, lingerie / topless / nude, same engine and control panel): https://civitai.red/models/2967745 . Follow my profile for more ComfyUI workflows.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚠️ Known limitations (please read)
In our testing: outfits described as opaque or lined can render sheer or see-through; skin can look glossy or oily; a hand can land near the groin; eyes can come out closed. These are model-level behaviours and the prompt wording does not fully fix them — review every image before you post it and discard anything that is off.
v1.3 only: in our comparison against v1.2 the skin highlights on legs and thighs looked a little stronger, and fine fabric texture (for example denim) looked smoother. We did not test other sizes, Upscale, Fix, BEAUTY=1 or a 12 GB card with v1.3.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📝 Notes & AI Disclosure
• All images are AI-generated with Qwen-Image 2.1 and show fictional adult people — no real person is depicted.
• Clothing guards are prompt-based: check every image before you post it. Discard anything unclear or cropped.
• Content rating for the showcase images: PG-13.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⭐ Found this useful?
• Like if it saved you time
• Comment your results — I read every one
• Follow for new ComfyUI workflows, all tested on 16 GB VRAM
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚖️ Model Attribution & Licensing
Built with Qwen. Qwen-Image 2.1 by the Qwen team (Alibaba) — model files from Comfy-Org/Qwen-Image-2.1 on Hugging Face; please read the licence on that repository. Prompt expander: Qwen3-VL 8B (same repository).
• v1.3 only: Viggle Turbo v0.3 by Viggle — https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo . It is a derivative of Qwen-Image 2.1 and is distributed by Viggle under the Qwen Research License (its LICENSE file has the same text as the base model's); read the LICENSE and NOTICE files in that repository before you use it. No Viggle file is included in this workflow, you download it from the source above.
• SeedVR2 (upscale): ByteDance-Seed, Apache-2.0; ComfyUI node by numz
• Node packs: WAS Node Suite (MIT), ComfyUI-Impact-Pack (GPL-3.0), rgthree-comfy (MIT), ComfyUI_essentials (MIT), ComfyUI-Image-Saver (MIT)
• Optional, not included: Qwen Image 2.1 Fix v1.0 by _Envy_ (Civitai) and our Qwen-Image 2.1 Glamour Realism LoRA v1.0 (Civitai)
Workflow JSON: CC0 Public Domain (free to reuse, no credit required).
🔗 Workflow by Thinni — feedback welcome!