Reference only

Qwen Image 2.1 downloads: component sets, quantization and source identity

Select the diffusion model, conditioning encoder and VAE separately. Keep optional prompt enhancers and file-size estimates out of minimum-VRAM claims.

Select a complete set

The reviewed official template selects the three files below. These are model-folder locations, not packages to install with pip. Q05

ComfyUI/models/
  diffusion_models/qwen_image_2.1_int8_convrot.safetensors
  text_encoders/qwen3vl_8b_int8_convrot.safetensors
  vae/qwen_image_2.1_vae_bf16.safetensors

Choose variants deliberately

The Comfy-Org card also lists a BF16 diffusion model and BF16 and W4A8 encoder variants. Q08 Treat the generator and conditioning encoder as separate selections. Match each selection to the loader and implementation you are using; a shared safetensors extension does not prove that a different quantization layout is supported. Keep the VAE in the set.

Do not substitute a prompt enhancer for the encoder

Files containing pe_t2i or pe_i2i are extra prompt-rewriting models. They do not replace qwen3vl_8b_*. The baseline graph does not become incomplete merely because an optional rewriter has not been downloaded. The publisher's pinned metadata provides SHA-256 values and exact byte counts for both enhancers at one repository commit; this site has not downloaded or hashed either file. Q14 Q15

Keep the source identity

Open the original file page and record the repository, relative path, revision, displayed size and license. The official template also links the Comfy-Org repository on ModelScope. That is an author-declared alternative entry point; this review has not checked file-by-file equality across the two hosts. Q05 Do not replace it with an unrelated mirror merely because the article is being read in Chinese.

Disk size is not a GPU budget

The W4A8 encoder page displays roughly 6.31 GB and the VAE page roughly 676 MB. Neither number measures the complete runtime. Q12 Q13 The publisher's pinned file metadata reports exact byte counts and LFS SHA-256 values for all eight listed files; these are source metadata, not local hash or GPU-memory measurements.

A filename can outlive its contents

The BF16 file history includes an old-file rename followed by an upload under the original name. Q16 Preserve revision and checksum information in addition to the filename. A checksum mismatch can mean that the upstream file changed, not only that a transfer failed. Compare identities before deleting a previously working copy. The eight inventory URLs now use an immutable repository commit; the model card and history views under main can change later.

Sources and verification

Source and pinned file metadata rechecked 2026-09-25 at repository commit 5dc5850eb514a3685f6a03a2641728a8f7549c69. No GPU run, weight download, local hash or quality benchmark. Procedures are editorial guidance, not a confirmed diagnosis of your environment.

Was this page helpful?

Anonymous. Only a yes/no counter is stored; no account, IP address or device details.

Sources & references

Official template and Comfy-Org file metadata rechecked 2026-09-25 at repository commit 5dc5850eb514a3685f6a03a2641728a8f7549c69. File hashes and byte counts are publisher-reported; no weight was downloaded or hashed locally, and no GPU run or quality benchmark was performed.

01Official Qwen 2.1 text-to-image templateSource checked: 2026-09-2502Comfy-Org Qwen 2.1 model cardSource checked: 2026-09-2503Qwen 2.1 VAE file metadataSource checked: 2026-09-2504Qwen3-VL W4A8 file metadataSource checked: 2026-09-2505Qwen PE-T2I Comfy file metadataSource checked: 2026-09-2506Qwen PE-I2I Comfy file metadataSource checked: 2026-09-2507Comfy BF16 file revision historySource checked: 2026-09-2508Pinned Qwen 2.1 file sizes and LFS SHA-256Source checked: 2026-09-25

Report an issue · 57974035-2378-59a4-95e6-9f24e4184a26