Qwen-Image 2.1 is the open image model my feeds won’t stop talking about this week. X posts mentioning it climbed to their highest daily count of the past week on October 24, Comfy-Org’s ComfyUI repack shows more than 10 million downloads in the last month on Hugging Face, and on October 22 Comfy-Org linked an official Qwen Image 2.1 Turbo text-to-image workflow from that repo. Turbo is the version that matters for home GPUs: the same 7B image model, distilled to 8 denoising steps at CFG 1, and Comfy-Org ships it as a 7.26 GB INT8 file. In this guide I explain which files to download for a 16 GB card like my RTX 5080, what every part of the official graph does, the settings you must leave alone, and how to get native 2K and transparent PNGs. You also get my 16 GB-friendly version of the workflow as a free download.
- Qwen-Image 2.1 Turbo (released October 9, 2026) is an 8-step distilled checkpoint of Qwen-Image 2.1, a 7B text-to-image and image-editing model with native 2K output and transparent (RGBA) image generation.
- For 16 GB GPUs, use Comfy-Org’s
qwen_image_2.1_turbo_int8_convrot.safetensors(7.26 GB), theqwen3vl_8b_int8_convrottext encoder (9.35 GB, or the 6.31 GBw4a8file) and the 0.68 GB VAE. - Keep the fixed 8-step
ManualSigmasschedule,eulerand CFG 1. The negative prompt does nothing at CFG 1. - Start at 1 MP (1376×768 at 16:9), then go to 2K (2720×1536 or 2048×2048) once a run works. Leave the 9B prompt enhancer off until then.
- Free download: my validated Qwen-Image 2.1 Turbo ComfyUI workflow (JSON + README with every model link) is in the download section. Core nodes only, no custom nodes.
What Is Qwen-Image 2.1 Turbo?
Qwen-Image 2.1 is the Qwen team’s unified text-to-image and image-editing model, released on September 20, 2026. According to the official GitHub README, its visual generation component has 7B parameters (32 single-stream DiT layers), and the release focuses on four things:
- Compact and efficient: mixed-granularity attention and prefix KV cache reuse.
- Native transparency: it can generate RGBA images from text, edit transparent layers and extract subjects from photos.
- Versatile editing: up to 10 reference images, local edits marked with circles, painted annotations or masks, and identity preservation for people and products.
- Better textures and typography: improved text rendering, portrait lighting and fine detail.
On October 9 the team released Qwen-Image-2.1-Turbo. The model card describes it as an accelerated checkpoint with the same 7B architecture that does text-to-image and editing in 8 denoising steps. It ships with its own sampling schedule and uses CFG 1 by default. The base model’s README default is 40 steps, so Turbo needs about a fifth of the steps, and with CFG 1 each step is a single model pass instead of two.
ComfyUI has supported Qwen-Image 2.1 natively since launch day, and Comfy-Org repackaged the weights in Comfy-Org/Qwen-Image-2.1, including INT8 “ConvRot” versions of the Turbo model and text encoder that are roughly half the size of the BF16 files.
Why It’s Trending Right Now
- X volume is rising: posts mentioning Qwen-Image 2.1 ran at roughly 70 to 135 a day from October 18 to 23, then hit about 200 on October 24, the highest day of the week (X recent-count API, checked October 25).
- Official ComfyUI workflows: on October 22 the Comfy-Org repo added links to the official Turbo text-to-image and Turbo image-edit templates.
- Downloads: the Comfy-Org repack shows over 10 million downloads last month on its Hugging Face page, and community quantized builds of 2.1 are on the Hugging Face trending list.
- Side-by-side tests: creators are comparing it with FIBO 2.0 and MOSS-Image, and in threads about what runs on 12 GB cards, “Qwen-Image 2.1 Turbo INT8” comes up as a recommendation.
What’s missing is a clear, 16 GB-focused guide to the Turbo workflow. That’s what this post is.
Files and Downloads for 16 GB GPUs
Everything comes from Comfy-Org/Qwen-Image-2.1. File sizes below are the ones Hugging Face lists (decimal GB).
Which Diffusion File: INT8 or BF16?
| File | Size | Use it when |
|---|---|---|
qwen_image_2.1_turbo_int8_convrot.safetensors |
7.26 GB | Default for 16 GB cards. It’s what the official template uses. |
qwen_image_2.1_turbo_bf16.safetensors |
14.23 GB | 24 GB+ cards, or if you want the unquantized weights. |
qwen_image_2.1_int8_convrot.safetensors / _bf16 |
7.26 / 14.23 GB | The base (non-Turbo) model, for 40-50 step runs or LoRA training. |
The INT8 Turbo file is about half the size of BF16, so it fits in 16 GB with room left over for activations at 2K. ComfyUI loads INT8 natively, so no GGUF loader or extra node pack is needed. One of ComfyUI’s own warning messages says its native formats (fp8, int8, w4a8) will be faster than GGUF even when they’re bigger than your memory.
Text Encoder, Prompt Enhancer and VAE
| Folder | File | Size | Notes |
|---|---|---|---|
| text_encoders | qwen3vl_8b_int8_convrot.safetensors |
9.35 GB | Default. Loaded with CLIPLoader type qwen_image. |
| text_encoders | qwen3vl_8b_w4a8.safetensors |
6.31 GB | Smaller alternative if system RAM or VRAM is tight. |
| text_encoders | qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors |
9.47 GB | Prompt enhancer. Optional, only used when refine_prompt is on. |
| text_encoders | qwen3.5_9b_qwen_image_2.1_pe_t2i_w4a8.safetensors |
6.72 GB | Smaller prompt enhancer. |
| vae | qwen_image_2.1_vae_bf16.safetensors |
0.68 GB | Required. |
Without the enhancer, the minimum download is about 17.3 GB (Turbo INT8 + INT8 text encoder + VAE). Add the INT8 enhancer and it’s about 26.8 GB. These files don’t all sit in VRAM together: the text encoder runs first, then the diffusion model samples, and ComfyUI’s memory manager moves models between VRAM and system RAM as needed. That’s why system RAM matters as much as VRAM here. On a 16 GB card I’d want 32 GB of system RAM or more.
Folder Layout
ComfyUI/
└── models/
├── diffusion_models/
│ └── qwen_image_2.1_turbo_int8_convrot.safetensors
├── text_encoders/
│ ├── qwen3vl_8b_int8_convrot.safetensors
│ └── qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors (optional)
└── vae/
└── qwen_image_2.1_vae_bf16.safetensorsUpdate ComfyUI first (official update guide). Qwen-Image 2.1 support landed in ComfyUI on September 19-20, and fixes followed through late September, including one for the int8/int4 cache option. I checked every node in my workflow against a ComfyUI 0.38.1 install, so v0.38 or newer is the safe minimum.
The Workflow, Node by Node
The official template, image_qwen_image_2_1_turbo_t2i.json in Comfy-Org’s workflow_templates repo, packs the pipeline into one subgraph node. That keeps the canvas tidy, but the settings that matter are hidden inside it. Here’s what’s in there.
Outer Graph: Resolution, Subgraph and Save
- Resolution Selector computes width and height from an aspect ratio and a megapixel target (1.0 MP = 1024 x 1024 pixels), rounded to a multiple you choose. The template’s size note says to prefer multiples of 32, so I set the multiple to 32.
- Text to Image (Qwen Image 2.1) subgraph shows the controls you’ll actually use: positive prompt,
refine_prompt,thinking_mode, sampler, the four model pickers,cache_deviceanddtype. - Save Image (Advanced) writes PNG, 8-bit, sRGB. PNG matters if you generate transparent images, because it keeps the alpha channel.
Inside the Subgraph: Loaders, Encoder and Cache
- Load Diffusion Model (UNETLoader) loads the Turbo INT8 file with weight dtype
default. - Load CLIP loads
qwen3vl_8b_int8_convrotwith typeqwen_image. A wrong type here is the most common reason this kind of workflow fails. - Text Encode Qwen Image 2.1 encodes both the positive and negative prompt. It can also take up to 16 reference images, which the edit template uses, but for text-to-image nothing is connected there.
- Qwen Image 2.1 Cache is an experimental core node that controls where the model’s prefix KV cache lives.
autouses spare VRAM, then RAM.cpukeeps it in RAM and, according to the node’s tooltip, is prefetched behind compute so it costs little speed.offrecomputes it every step. Thedtypeoption can store the cache asint8(half the size, about bf16 accuracy) orint4(a quarter of the size, but roughly double the per-step error). - Empty Latent Image gets its width and height from the Resolution Selector.
Sampling: ManualSigmas, Euler and CFG 1
Sampling uses ComfyUI’s custom-sampler nodes instead of a plain KSampler:
- ManualSigmas:
1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0. Those nine values define the eight steps. This is the schedule Turbo was distilled for, and the Turbo model card says other schedules haven’t been evaluated. - KSamplerSelect:
euler. - CFGGuider at 1.0, fed by the cache-patched model and both conditionings.
- RandomNoise and SamplerCustomAdvanced, then VAE Decode with the Qwen-Image 2.1 VAE.
Look closely at the schedule: most steps are spent at high noise (1.0 down to about 0.85), then it drops fast. Distilled models are trained on a few specific noise levels, so the generic simple or karras schedulers would ask the model to denoise at levels it never learned. That’s why the template uses ManualSigmas instead of a scheduler.
The Optional Prompt Enhancer
The top half of the subgraph is a full prompt-rewriting pipeline: a second CLIPLoader loads the Qwen 3.5 9B enhancer, a Generate Text (TextGenerate) node rewrites your prompt using a long system prompt stored in a PrimitiveStringMultiline node, Preview as Text shows the result, and a Switch node picks either your prompt or the rewritten one. The Qwen team recommends their rewriting models for best results, since the model was trained on long, detailed descriptions.
The template note gives the settings: turn thinking_mode on together with refine_prompt (the enhancer was trained with a reasoning block and loses quality without it), and max_length is capped at 4096 tokens instead of the official 16,256 so one rewrite can’t run for many minutes. Sampling values follow the official release: temperature 1.0, top_p 0.95, top_k 20, presence penalty 1.5.
On a 16 GB card, the enhancer means loading another 9B model before the image model, so I leave it off at first. My workaround is to write long, specific prompts by hand, or turn the enhancer on once, copy the rewrite from the preview node, and then switch it off.
Settings You Must Not Change
| Setting | Value | Why |
|---|---|---|
| Steps / schedule | 8 via ManualSigmas (values above) | The schedule Turbo was distilled for. To use more steps, you’d also need a different schedule. |
| Sampler | euler | Template default for the distilled path. |
| CFG | 1 | Turbo’s default. Above 1, every step needs two passes, and the template says to raise it only if you actually use a negative prompt. |
| CLIP type | qwen_image | Wrong type = wrong tokenizer/template = broken images. |
16 GB VRAM Tips
- Get one 1 MP image first. My workflow starts at 1376×768 (16:9, 1 MP). Once that works, go up in size.
- Set
cache_devicetocpuon the subgraph if you hit out-of-memory errors at 2K. The node’s own description says the RAM cache costs little speed. - Swap to the w4a8 text encoder (6.31 GB) if loading the encoder makes your system swap. You only need to change the
clip_namepicker. - Keep the prompt enhancer off until the base run is stable. When you use it, pick the w4a8 enhancer (6.72 GB).
- Don’t use
int4cache dtype unless you have to: the node’s tooltip says it roughly doubles the per-step error. Tryint8first. - Close other GPU apps. On my PC, ComfyUI shares the card with other tools, and anything else holding VRAM slows every model.
I haven’t benchmarked this exact graph on my RTX 5080 yet, so this post has no timing numbers. I’ll add measured times once I have them, not before.
Resolutions: 1 MP, 2 MP and Native 2K
Qwen-Image 2.1 natively supports 2K output, and both checkpoints default to the 2048 resolution level in Diffusers. Using Resolution Selector with multiple 32:
| Aspect | Megapixels | Output | Use |
|---|---|---|---|
| 16:9 | 1.0 | 1376 x 768 | First runs, quick iteration (workflow default) |
| 16:9 | 2.0 | 1920 x 1088 | Blog headers, YouTube thumbnails |
| 16:9 | 4.0 | 2720 x 1536 | Close to the official 2752 x 1536 preset |
| 1:1 | 4.0 | 2048 x 2048 | The native 2K square size from the template note |
| 9:16 | 2.0 | 1088 x 1920 | Shorts and Reels covers |
If you want the exact official presets (for example 2752 x 1536 for 16:9 or 1696 x 2528 for 2:3), disconnect the Resolution Selector and type the width and height into the subgraph directly. They’re all multiples of 32.
Transparent PNGs with Qwen-Image 2.1
This is the feature I’m most excited about for creator work: stickers, thumbnail cut-outs and overlay graphics without a background-removal pass. The official README recommends this prompt format:
This is an RGBA image with transparency. A cute cartoon dragon sticker. The image has alpha channel and the background is transparent.Replace the middle sentence with your subject and keep the output as PNG (the Save Image node in my workflow already is). The Comfy-Org template note uses almost the same wording. Comfy-Org also has a separate “Remove Background: Qwen Image 2.1” template if you need to cut a subject out of an existing photo.
Prompting Tips for Turbo
- Write like you’re describing a finished photo. The official examples are long, concrete paragraphs: subject, setting, lighting, lens, materials and layout. Short tag lists waste the model’s strengths.
- Put text in quotes. For signs and posters, write the exact words in double quotes and say where they go (for example: a wooden signboard above the stall that reads “JOLLY CHAI” in bold yellow letters). Keep on-image text short at 1 MP and go to 2K for dense typography.
- Say what you don’t want inside the positive prompt. At CFG 1 the negative prompt is ignored, so phrases like “no watermark” or “spelling exactly correct” belong in the main prompt, as the official template’s own example does.
- Fix the seed while tuning. My workflow ships with a fixed seed so you can compare prompt changes one at a time.
Turbo vs Base Model vs Turbo LoRA
| Option | Steps | CFG | Best for |
|---|---|---|---|
| Turbo checkpoint (this workflow) | 8 (fixed schedule) | 1 | Fast everyday generation on 16 GB cards |
| Base Qwen-Image 2.1 | about 40-50 (euler) | above 1, with negative prompt | Maximum control, negative prompts, LoRA training base |
Base + Turbo LoRA (qwen_image_2.1_turbo_lora_avg_rank_178_bf16, 0.91 GB) |
8 | 1 | You already keep the base model and don’t want a second 7 GB checkpoint |
The Comfy-Org model card offers both: “use either the full Turbo model, or the Turbo LoRA extracted from the base model. Use 8 steps and cfg 1.” One thing to watch with community LoRAs: at least one popular anime LoRA author reported quality loss when applying a base-model LoRA to Turbo and has since put out an experimental Turbo-specific version. So check which checkpoint a LoRA was trained on before you blame the workflow.
Download the Qwen-Image 2.1 Turbo ComfyUI Workflow
The zip contains Qwen_Image_2.1_Turbo_T2I_16GB.json (ComfyUI UI format, drag it onto the canvas) and a README.txt listing every model file, its download link and its folder. It’s built on Comfy-Org’s official Turbo text-to-image template, and the 8-step sampler graph is unchanged. My changes:
- 16:9 at 1 MP with multiple 32 (1376×768) instead of 1:1 with multiple 8
- A fixed seed (20261025) and a photo prompt with one short sign, so a single run shows both photorealism and text rendering
- Output saved to
ComfyUI/output/FreakyJolly/ - An extra note on the canvas with the 16 GB swaps, 2K sizes and the RGBA prompt format
Before publishing, I validated it with a script: valid JSON, every link (in the main graph and inside the subgraph) pointing at existing nodes and slots with matching types, and every node class checked against current ComfyUI source and a live ComfyUI 0.38.1 node list. No custom nodes needed. Model weights aren’t included (Qwen Research License).
Download Qwen-Image 2.1 Turbo ComfyUI Workflow (.zip)
Requirements: ComfyUI v0.38 or newer with a current frontend (the workflow uses a subgraph), an NVIDIA GPU with 16 GB VRAM recommended, about 17.3 GB of model downloads (26.8 GB with the prompt enhancer), and 32 GB+ system RAM recommended.
Frequently Asked Questions
Will it run on 12 GB or 8 GB cards?
The INT8 Turbo file is 7.26 GB, so 12 GB is realistic at 1 MP with cache_device set to cpu and the w4a8 text encoder. On 8 GB, ComfyUI will offload heavily and it will be slow. Community GGUF builds exist, but ComfyUI recommends its native formats over GGUF.
Can I use 12 or 20 steps for more detail?
Not by just changing a step count. The eight steps come from the ManualSigmas list, and the Turbo card says other schedules haven’t been evaluated. For more steps, use the base model at 40-50 steps.
Why doesn’t my negative prompt work?
At CFG 1 the negative conditioning has no effect. Put exclusions in the positive prompt, or use the base model with CFG above 1.
Can this workflow edit images?
This one is text-to-image only. Turbo also edits images; use the official “Qwen Image 2.1 Turbo: Image edit” template, which connects reference images to the same Text Encode Qwen Image 2.1 node.
ComfyUI shows red or missing nodes. What now?
Update ComfyUI and its frontend. All nodes here are core nodes. If the subgraph looks empty or broken, your frontend is too old to read subgraphs.
Can I use the images commercially?
Read the license first. The open weights use the Qwen Research License Agreement, which grants rights for non-commercial purposes only and asks you to contact Qwen for a commercial license. If you need commercial use, Qwen also offers hosted Qwen-Image 2.1 Pro and Turbo APIs on Alibaba Cloud Model Studio.
Sources
- QwenLM/Qwen-Image-2.1 on GitHub (architecture, news, RGBA prompt, aspect ratios, prompt rewriting)
- Qwen/Qwen-Image-2.1-Turbo model card (8 steps, CFG 1, saved schedule)
- Comfy-Org/Qwen-Image-2.1 (ComfyUI files, sizes, Turbo LoRA note)
- Comfy-Org official template image_qwen_image_2_1_turbo_t2i.json
- ComfyUI nodes_qwen.py (Text Encode Qwen Image 2.1, Qwen Image 2.1 Cache)
- Qwen/Qwen-Image-2.1-PE-T2I (prompt enhancer)
- Qwen-Image 2.1 blog
