CivArchive
    Qwen Image 2.1 12GB — int8 Native 2K Text to Image, 40-step QUALITY + Viggle 6-step TURBO LoRA - v2.0.1
    Preview 149090280
    Preview 149090279
    Preview 149090283
    Preview 149090281
    Preview 149090354
    Preview 149090355
    Preview 149090356
    Preview 149090357

    ☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1


    Native 2K Qwen Image 2.1 on a 12 GB card, with two paths in one graph. A QUALITY path at 40 steps with Comfy Kitchen attention, and a one-click TURBO path on Viggle's 6-step LoRA. Built from the official Text to Image template and retuned from my own benchmark: int8 model, int8 text encoder, 2048×2048, 40 steps instead of the template's 25.

    What you need

    ✔ Custom nodes — None. Core ComfyUI only.
    ✔ ComfyUI version — 0.37 or newer — native Qwen Image 2.1, ModelAttentionBackend, QwenImage21Cache. 0.37.1 tested.
    ✔ VRAM — 12 GB — tested, not estimated
    ✔ System RAM — 3.0 GB peak on int8 (bf16 offloads and peaks at 10.6 GB)
    ✔ Tested on — RTX 3060 12 GB · 32 GB DDR4 · ComfyUI 0.37.1 (Comfy Desktop)
    ✔ Licence — Workflow MIT · Qwen Image 2.1 and the Viggle LoRA: Qwen Research License (non-commercial)

    Models to download

    Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/.

    Required — 3 files, 17.38 GB

    diffusion_models/

    File: qwen_image_2.1_int8_convrot.safetensors — 7.3 GB (Diffusion)
    Download:
    https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/diffusion_models/qwen_image_2.1_int8_convrot.safetensors

    Put it here:
    ComfyUI/models/diffusion_models/

    text_encoders/

    File: qwen3vl_8b_int8_convrot.safetensors — 9.4 GB (Text encoder)
    Download:
    https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/text_encoders/qwen3vl_8b_int8_convrot.safetensors

    Put it here:
    ComfyUI/models/text_encoders/

    vae/

    File: qwen_image_2.1_vae_bf16.safetensors — 0.68 GB (VAE)
    Download:
    https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors

    Put it here:
    ComfyUI/models/vae/

    TURBO path only — 1 file.

    loras/

    File: Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r256.safetensors — 1.4 GB (Turbo LoRA)
    Download:
    https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/resolve/main/Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r256.safetensors

    Put it here:
    ComfyUI/models/loras/

    Optional swaps, same repos, not in the workflow JSON: qwen_image_2.1_bf16.safetensors (14.2 GB), the bigger diffusion model, about 1.6× slower on 12 GB because it offloads to RAM · qwen3vl_8b_w4a8.safetensors (6.3 GB), smaller text encoder, same speed, different picture on the same seed.

    Speed on my RTX 3060

    This exact workflow, Comfy Desktop with SageAttention on, 2048×2048, warm run:

    • QUALITY · 40 steps · Comfy Kitchen — 4 min 07 s time per image

    • TURBO · Viggle 6-step — 47 s time per image

    From the full benchmark, same machine, 2048×2048, 25 steps, same seed: PyTorch attention 3 min 57 s · SageAttention 3 min 00 s · Comfy Kitchen 2 min 40 s · Viggle 6-step 62 s on PyTorch and 50 s on Sage · Viggle 4-step 44 s.

    int8 against bf16, 2048×2048, 25 steps, PyTorch attention: 3 min 57 s against 6 min 10 s, model file 7.3 GB against 14.2 GB, peak system RAM 3.0 GB against 10.6 GB. bf16 does not fit in 12 GB, it offloads. Same-seed int8 and bf16 images measured SSIM 0.89–0.99 over 4 prompts.

    Known issues

    • TURBO is softer in fine detail at 2K than the base model. Measured, and visible in 100% crops.

    • Seeds vary. Of two test seeds, TURBO double-exposed the subject's head on one and QUALITY printed the date line twice on one. Re-roll the seed.

    • Do not add Comfy Kitchen attention to TURBO. It ghosted the subject with the 6-step LoRA in my test. Leave TURBO on the ComfyUI default, which is Sage on Comfy Desktop.

    • TURBO sigmas depend on the size. Change the size, change the line — the table is in the workflow note.

    • The Viggle LoRA goes through ComfyUI's standard merging LoRA loader. Viggle's own unmerged-LoRA node was not used.

    • No peak VRAM figure was measured for this one, only peak system RAM. So the 12 GB number is "it runs on a 12 GB card", not a nvidia-smi reading.

    Get the workflow

    Second video, the settings one: https://www.youtube.com/watch?v=-yJvQ3lxWsE&list=PLG_da2oqTx7k

    More videos on this stuff: https://www.youtube.com/playlist?list=PLG_da2oqTx7k


    Description

    v2.0.1: Notes only, graph unchanged. "How to use" note rewritten: what it does, each input on the main node, outputs, how to switch to TURBO (incl. enabling the bottom Save Image), important settings, tips, the attention flags line, and a Tested on list (RTX 3060 + ComfyUI 0.45.3 on a 24 GB card). Downloads note: punctuation only, text unchanged.

    Workflows
    Qwen 2.1

    Details

    Downloads
    29
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/15/2026
    Updated
    11/17/2026
    Deleted
    -

    Files

    The_frizzy1_qwen-image-2.1-12gb_v2.0.1.json