Zhenzhen's AI Radar · Qwen Image 2.1 · 2026-11-14 Morning Report
This issue features 6 new versions: Strongly Recommended 1 · Recommended 1 · Worth Watching 2 · Included Only 2
Editor's Note This issue reviews 6 new versions of Qwen 2.1: 1 rated S, 1 rated A, 2 rated B, and 2 rated C. The top recommendation is The_frizzy1's Ultimate 12GB workflow: a single dropdown menu switches between 2K text-to-image, editing 1–3 images, background removal to transparent PNG, and Fun ControlNet pose/depth/line art. The author even documented failure cases. The same author's 12GB text-to-image workflow has detailed speed tests but overlaps with the text-to-image path of the top recommendation, so it is rated A; Detailifier's Qwen 2.1 port lacks version notes, and the parameters on the card belong to the Flux version, so its automatic S rating is downgraded to B. The eerie surreal style LoRA has only a one-sentence description and is rated B; two realistic/original character LoRAs are rated C, with no exclusions.
Strongly Recommended (1)
[S] Qwen Image 2.1 Ultimate 12GB — Turbo 8-step 2K T2I + Edit 1-3 Images + Remove Background + Fun ControlNet Pose/Depth/Canny · v2.1.0
Workflow · New Model · Author The_frizzy1 · Published 11-14 05:40 · Downloads 14
A workflow that consolidates multiple modes of Qwen Image 2.1 into a single dropdown menu: 2K text-to-image, editing 1–3 images, background removal directly to transparent PNG, and Fun ControlNet pose/depth/line art, optimized for 12 GB GPUs.
Images

Reasons for Recommendation
- New workflow, defaults to the official Turbo model with 8 steps; uses only core nodes, requires ComfyUI 0.38 or higher
- v2.1.0 adds a background removal mode: follows the editing path and fixed prompts of the official 0.39 template, preserving the alpha channel; the author claims 1 MP takes 29 seconds on Turbo and 54 seconds on the base model
- The author tested on an RTX 3060 with 12 GB, reporting peak VRAM usage of 11.6 GB (Turbo, 2K); 12 visible examples
How to Use
- Select the mode via the Mode dropdown; keep unused image inputs bypassed (purple), otherwise you will get an Invalid image file error before running
- In edit mode, use <Picture 1> etc. in the prompt to refer to images 1–3; turn off Turbo to switch to the base model with 40 steps
- ControlNet mode additionally requires files such as Fun ControlNet union, SDPose, and Lotus; keep resolution at 1 MP
Risks and Limitations
- Author's note: ControlNet breaks at native 2K (pose backgrounds produce chaotic lines, depth outputs raw depth maps), so only 1 MP is usable; background removal results in almost fully transparent images for large indoor scenes after 3 attempts
- Loaders for un-downloaded models will be marked red, requiring you to point to any existing file; 5 examples are hidden
- Qwen Image 2.1 and Turbo are under the Qwen Research License, non-commercial use only
Recommended (1)
[A] Qwen Image 2.1 12GB — int8 Native 2K Text to Image, 40-step QUALITY + Viggle 6-step TURBO LoRA · v2.0.0
Workflow · New Model · Author The_frizzy1 · Published 11-14 05:40 · Downloads 12
A workflow for running Qwen Image 2.1 native 2048×2048 text-to-image on 12 GB GPUs: the same image contains a 40-step QUALITY path and a Viggle 6-step TURBO path.
Images

Reasons for Recommendation
- New workflow, based on the official text-to-image template; the author adjusted the step count from the template's 25 steps to 40 steps based on their own speed tests, using Comfy Kitchen attention for the QUALITY path
- Uses only core nodes, requires ComfyUI 0.37 or higher; the author claims the int8 version has a peak system RAM usage of 3.0 GB, while bf16 offloads to RAM with a peak of 10.6 GB
- 13 visible examples
How to Use
- Requires 3 files: int8 main model, int8 text encoder, VAE; the TURBO path additionally requires the Viggle 6-step LoRA (v0.2.1, r256)
- TURBO sigmas vary with size; when changing size, update them according to the table in the workflow notes; keep TURBO's default attention, do not switch to Comfy Kitchen
Risks and Limitations
- Author's note: TURBO details are softer than the base model at 2K; combining TURBO with Comfy Kitchen attention causes ghosting; test seeds showed head ghosting and repeated date lines
- The author claims QUALITY takes about 4 minutes 07 seconds per image and TURBO about 47 seconds on RTX 3060; peak VRAM was not tested, so 12 GB only indicates it runs
- The version note is an invalid auto-generated text, not stating what changed in v2.0.0; overlaps with the text-to-image path of the Ultimate workflow published by the same author at the same time; Qwen Research License is non-commercial
Worth Watching (2)
[B] Detailifier · v1.0 Q2.1
LoRA · Port · Author zwaetschge · Published 11-13 23:59 · Downloads 103
The Qwen 2.1 version of the detail enhancement LoRA: the author claims it enhances facial, skin, clothing, hair, and material details, usable in both text-to-image and image-to-image/upscaling workflows.
Images

Reasons for Recommendation
- Ported version: the same model already has versions for SDXL, Pony, Flux.1 D, Krea 2, MOSS, etc.; this release ports it to Qwen 2.1
- 3 visible examples; about 9 hours after release, 103 downloads, with attention mainly stemming from the legacy model's accumulation
How to Use
- Author's general advice: strength 0.2–1, use at 2–4 MP resolution; no trigger word
- The CFG 2.0–2.5 and 20–30 steps on the card are recommendations for the Flux version; the author did not provide parameters for the Qwen 2.1 version
Risks and Limitations
- The Qwen 2.1 version lacks version notes, and differences from other base model versions are unknown
- 2 examples are hidden; the license only allows on-site rental
[B] Creepy Surreal · v1.0 Q2.1
LoRA · New Model · Author zwaetschge · Published 11-14 04:43 · Downloads 17
Eerie surreal style LoRA: the author claims it was trained on 35 eerie doll-like surreal images, serving horror and grotesque illustration themes.
Images

Reasons for Recommendation
- New style LoRA with a trigger word; 5 visible examples allow judging the art style
How to Use
- Add creepysurreal to the prompt; the author did not provide strength or step count
Risks and Limitations
- Description is only one sentence, and the training set is small
- 1 examples are hidden; the license does not allow publishing derivative models
Trigger Word: creepysurreal
Included Only (2)
- Kira — Photorealistic LoRA [MOSS / Krea 2,Qwen 2.1] · v2.0 Qwen 2.1 — LoRA: Qwen 2.1 v2 of the realistic female character LoRA "Kira": the author claims v2 in other base model versions is softer, warmer, and more everyday.

- FEKLA [MOSS / Krea] · v1.0 Qwen 2.1 — LoRA: Qwen 2.1 version of the LoRA for the AI-generated original character (OC) "FEKLA": the author claims it fixes multiple details such as colorful tattoos, pearl chokers, and two-tone bangs onto the same character.

Data: Civitai public API (base model Qwen 2.1), inclusion window 11-13 20:31 → 11-14 09:18 (Beijing Time). Recommendation reasons have been reviewed by editors; download counts only reflect attention.