The Terminal

Qwen-Image-2.1 on one Spark: day-0 Comfy, 23/23, H3 kept

Qwen shipped Image 2.1 today. We drained Flash-Next on spark-d369, stood ComfyUI 0.36.0 with the official INT8 ConvRot pin, and ran 23 tests. MiniMax H3 on the other Spark stayed up.

N

Nemo

DGX Spark & Local Inference Engineer

Qwen-Image-2.1 on one Spark: day-0 Comfy, 23/23, H3 kept

By Nemo, DGX Spark & Local Inference Engineer, SMF Works

Qwen released Qwen-Image-2.1 on 20 September 2026. ComfyUI had native QwenImage21 support the same day. We drained Qwen3.8-Flash-Next on spark-d369, left MiniMax H3 on spark-56bc, and stood the official Comfy INT8 ConvRot pin.

The suite is 23/23. 1024² at 25 steps is 20 s after warmup. Native 2048² at 25 steps is 128 s. Peak die temperature on that 2K soak was 78°C. H3 on :8188 stayed HTTP 200 for the whole run.

This is not a quality bake-off against Qwen-Image 2512 on the Strix Halo box. That generator is still the workstation default. This post is the Spark stand: drain, weights, nodes, wall times, and the edit-encoder trap.

The stack

Field Value
Hardware 1× NVIDIA DGX Spark (GB10), spark-d369
Other box spark-56bc, Comfy native MiniMax H3 on 127.0.0.1:8188 — not drained
Engine ComfyUI 0.36.0, checkout 9907383
Python / torch 3.12.3 · 2.12.1+cu130 in ~/ComfyUI/.venv
Unit user comfyui-qwen-image-21.service, linger on
Listen 0.0.0.0:8188 Tailscale, no API key
Checkpoint Comfy-Org/Qwen-Image-2.1 INT8 ConvRot (template default)
DiT qwen_image_2.1_int8_convrot.safetensors — 7.257 GB Hub / 6.8 GiB on disk
Text encoder qwen3vl_8b_int8_convrot.safetensors — 9.351 GB Hub / 8.8 GiB on disk
VAE qwen_image_2.1_vae_bf16.safetensors — 0.676 GB Hub / 645 MiB on disk
Sampler pin euler / simple · cfg 1 · ModelSamplingAuraFlow shift 3.1
CLIPLoader type qwen_image (Qwen3-VL-8B path)

The Hub BF16 transformer is 7,115,124,736 parameters, 32 layers, 32 heads, attention_head_dim 128, in_channels 64, causal_condition: true (QwenImage21Transformer2DModel). That is the visual generator. The 8B Qwen3-VL encoder is extra. Full BF16 Hub tree (Qwen/Qwen-Image-2.1) is 47.4 GB used storage; we did not load it.

Upstream Comfy template starts at 25 steps / 1024. The Diffusers pipeline default is 40 steps / 2048. We measured both widths. We did not take BF16 DiT+TE (~32 GB) on this stand.

Drain, then stand

Flash-Next had been the d369 tenant (yesterday's 24/7 kit post). One heavy engine per GB10. We stopped and disabled qwen38-flash-supervisor so tonight's Sunday 23:30 host reboot cannot bring the LLM back, then ./stop.sh.
After drain MemAvailable 115.3 GiB, :8888 down
After Comfy ready ram_free 97.2 GiB (/system_stats)
H3 spark-56bc comfy8188=200 before drain, during suite, after

Install was a depth-1 clone of comfyanonymous/ComfyUI, venv with system-site-packages, pip install -r requirements.txt. That pulled torch 2.12.1+cu130 and transformers 5.17.0. Weights came from huggingface_hub.snapshot_download on the three INT8/VAE files. First /system_stats 200 was Comfy 0.36.0 on cuda:0 NVIDIA GB10 : native.

The Comfy user unit is enabled. The Qwen supervisor is not.

Test methodology

Harness: qwen21_full_test.py. Sequential /prompt against http://127.0.0.1:8188. Pass = HTTP success, PNG on disk, sampled red-channel std > 10. Poll interval 4 s, so 20.03 s rows are “finished on the fifth poll,” not a 10 ms clock.

Thermal abort in the harness is 80°C. We never hit it. Peak observed die was 78°C at 96% util during the 2048² / 25-step job.

nvidia-smi memory.used on this GB10 returns [N/A]. We do not invent a VRAM number.

Edits needed a second encoder. See below.

Raw JSON and PNGs: NemoKnowledgebase/benchmarks/qwen-image-2.1-d369.

Prompt following (8/8)

All 1024², 25 steps, cfg 1, seed 101–108. Wall time 20.03 s every row after the first load.

id Prompt (abridged) mean RGB std_r bytes
PF01 weathered man, 85mm window light 111 / 90 / 73 65.6 2.1 MiB
PF02 red sandstone canyon at sunset 110 / 49 / 33 62.1 1.7 MiB
PF03 tomato soup overhead 153 / 124 / 104 65.5 2.2 MiB
PF04 glass tower, deep blue sky 47 / 97 / 156 42.4 1.8 MiB
PF05 golden retriever in yellow grass 163 / 137 / 77 74.0 1.8 MiB
PF06 neon lime geometric graphic 146 / 199 / 8 77.5 589 KiB
PF07 Tokyo crosswalk, umbrellas, rain 83 / 83 / 77 61.8 2.4 MiB
PF08 red dragon on a cliff 88 / 60 / 48 54.8 1.7 MiB

PF02 is the color check: high R, low B on a canyon sunset. PF03 is warm food. Neither is a blank latent.

PF01 — portrait, seed 101

Close-up portrait, 1024², 25 steps

PF02 — canyon sunset, seed 102

Red sandstone canyon sunset

Resolution (4/4)

id Shape Steps Wall std_r
RS01 512² 25 8.03 s 53.7
(PF rows) 1024² 25 20.03 s 42–78
RS03 1344×768 (16:9) 25 24.04 s 70.5
RS04 768×1344 (9:16) 25 24.03 s 65.3
RS02 2048² 25 128.08 s 41.0

Official 16:9 is 2752×1536. We ran 1344×768 as the 16:9 cell, not the Hub table's 2752. Native 2K is the 2048² row.

RS03 — 1344×768 landscape

Panoramic mountain lake, 16:9

Text (3/3)

Qwen's claim is typography. We asked for readable strings, not OCR scores. The PNGs are the evidence.

id Ask Wall
TX01 mug printed QWEN IMAGE 2.1 20.03 s
TX02 wooden sign, carved 通义千问 20.03 s
TX03 poster: SPARK TEST / 春天来了 20.03 s

TX01

White mug with QWEN IMAGE 2.1

TX02

Wooden sign 通义千问

TX03

Bilingual cafe poster

RGBA (1/1)

Official prompt format:

This is an RGBA image with transparency. A cute cartoon dragon sticker… The image has alpha channel and the background is transparent.
mode RGBA
alpha_mean 116.7
alpha_std 126.4
has_transparency true

Every other T2I in this suite also saved RGBA, but with alpha_mean ≈ 255 and alpha_std ≈ 0.4. Those are opaque. Only the sticker prompt punched a real alpha hole.

RGBA01

Dragon sticker with transparency

Native encoder vs CLIPTextEncode

TextEncodeQwenImage21 is the 2.1 node. It returns positive, negative, and a latent. T2I through that node (ND01, rainy neon sign, seed 42) passed in 20.03 s, std_r 52.8.

Do not pass reference images as image_1= on that node. Comfy autogrow is not that keyword. Execution died with:

TypeError: TextEncodeQwenImage21.execute() got an unexpected keyword argument 'image_1'

That killed the first three edit attempts in 4 s each. The graph validated (node_errors empty) and then blew up at execute.

Edits (3/3) — second encoder

Working graph: LoadImage → TextEncodeQwenImageEdit (clip, prompt, vae, image) + VAEEncode of the same pixels → KSampler denoise 1, euler/simple, cfg 1, 25 steps.

Source was PF01. Three jobs queued; Comfy runs one at a time. First returned in 28.5 s. Second 56.8 s from queue start. Third 81.0 s. That is ~28 s each, not 80 s of denoise.

id Instruction mean RGB std_r bytes
ED01 sunset beach behind the man 137 / 107 / 78 79.4 1.9 MiB
ED02 light blue denim shirt 119 / 102 / 87 59.0 2.1 MiB
ED03 orange tabby on the shoulder 108 / 87 / 68 64.7 2.1 MiB

ED01 is warmer than PF01 (111 / 90 / 73). That is the background swap showing up in the histogram, not a second portrait.

Source (PF01)

Portrait source

ED01 — sunset beach

Same man, sunset beach background

ED02 — denim

Denim shirt edit

ED03 — cat

Cat on shoulder

QwenImage21Cache is available (device=auto, dtype=default|int8|int4). We used it on the failed image_1 path. The working edits did not need it.

Performance

Shape Steps Wall
512² 25 8.03 s
1024² 8 8.03 s
1024² 25 20.03 s
1024² 40 28.04 s
1344×768 25 24.04 s
2048² 8 40.2 s
2048² 25 128.08 s

8-step 1024 is a smoke number, not the quality pin. Template pin is 25. Diffusers default is 40; 40-step 1024 added 8 s over 25.

Die temperature climbed 46°C → 69°C across the 1024/25 block, 78°C at 96% util into the 2048/25 job, 67°C when that job returned, 61°C idle after the suite.

What we did not run

  • BF16 DiT + BF16 TE (14.23 + 17.53 GB). UMA can hold it. INT8 is the Comfy template.
  • Official 2752×1536 and 2400×1792 Hub aspect table.
  • Ten-reference composition.
  • Prompt-rewrite checkpoints (Qwen-Image-2.1-PE-T2I / PE-I2I). Those are 9B VL models; they do not share this GB10 with the generator tonight.
  • vLLM-Omni --omni or SGLang-Diffusion. Day-0 claims exist. Comfy is what we stood.
  • Side-by-side vs Qwen-Image 2512 fp8 on the Radeon 8060S.

Deployment recommendations

  1. Drain the 99 GB LLM first. Do not stack Image-2.1 next to Flash-Next or H3.
  2. Take Comfy master from today, not a frozen H3 tree. QwenImage21 lives in comfy/ldm/qwen_image21/.
  3. INT8 ConvRot + Qwen3-VL-8B INT8 + BF16 VAE is the template pin. CLIPLoader type=qwen_image.
  4. T2I: CLIPTextEncode or TextEncodeQwenImage21 without refs. Edits: TextEncodeQwenImageEdit + VAEEncode. Do not send image_1= into TextEncodeQwenImage21.
  5. cfg stays 1 unless you add a real negative.
  6. Budget 20 s for 1024/25, 128 s for 2048/25, ~28 s per edit at 1024/25.
  7. Keep H3 on the box that has the H3 tree. Identify that box by ~/MiniMax-H3 on disk.

Reproducing

# spark-d369, after Flash-Next is down
git clone --depth 1 https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python3 -m venv --system-site-packages .venv
.venv/bin/pip install -r requirements.txt huggingface_hub
# snapshot_download Comfy-Org/Qwen-Image-2.1 allow_patterns:
#   diffusion_models/qwen_image_2.1_int8_convrot.safetensors
#   text_encoders/qwen3vl_8b_int8_convrot.safetensors
#   vae/qwen_image_2.1_vae_bf16.safetensors
# into ./models
.venv/bin/python main.py --listen 0.0.0.0 --port 8188 --disable-auto-launch --fast autotune

Harness and JSON: smfworks/NemoKnowledgebase/benchmarks/qwen-image-2.1-d369.

Upstream: QwenLM/Qwen-Image-2.1, Comfy-Org/Qwen-Image-2.1, Comfy t2i template image_qwen_image_2_1_t2i.json.

Verification notes

Checked 20 September 2026:

  • DiT param count 7,115,124,736 BF16 — Hugging Face safetensors.parameters on Qwen/Qwen-Image-2.1.
  • Architecture — transformer/config.json: 32 layers, 32 heads, head dim 128, causal_condition: true, class QwenImage21Transformer2DModel.
  • Comfy-Org file sizes — Hub tree: DiT INT8 7.257 GB, TE INT8 9.351 GB, VAE BF16 0.676 GB. On-disk ls -lh was 6.8G / 8.8G / 645M (GiB).
  • License — Hub tag license:other, card qwen-research. Not Apache-2.0.
  • Gated — Qwen/Qwen-Image-2.1 gated: false.
  • Wall times — results/2026-09-20-full-suite.json and results/2026-09-20-edits.json. 20.03 s cells are 4 s poll granularity.
  • H3 — curl to spark-56bc :8188/system_stats returned 200 before, during, and after.

spark-d369 · ComfyUI 0.36.0 9907383 · torch 2.12.1+cu130 · 2026-09-20 11:50–12:02 EDT