Introduction

ComfyUI just hit v0.37.0, and the headline most of the timeline already shouted is true: day-zero Qwen-Image-2.1 templates are in core. We covered that open-weight drop here.

The quieter part of the same release is what I want on your update checklist tonight: MoGe 3 geometry estimation lands natively, Qwen text encoders get a serious speed pass, and Comfy finally auto-detects a fast NVMe and turns on disk-backed loading without a ritual flag.

If you live in node graphs for depth, normals, camera match, or just hate waiting on text encoders — this build is more than a Qwen party favor.

What Landed in v0.37.0

Published on GitHub as Comfy-Org/ComfyUI v0.37.0 (Sep 21, 2026). The creator-facing cluster:

  • Qwen-Image-2.1 in core — three official templates (T2I, edit, background removal), plus TextEncodeQwenImage21 and QwenImage21Cache for reference-image edits and KV-cache control. (Already the star of our Qwen-Image-2.1 writeup.)
  • MoGe 3 — Microsoft Research’s fine-detail monocular geometry model, wired into core with ViT-L / ViT-G checkpoints and a sparse volumetric refiner.
  • Faster Qwen3.5 / Qwen3.8 text encoders — speculative decoding (MTP), FixedKVBias attention cache, fused DeltaNet kernels, W4A8 GEMV path.
  • Auto fast-disk — comfy-aimdo 0.5.5 detects PCIe 4 NVMe (and friends) and enables disk-backed dynamic load/offload per model; --disable-fast-disk if you need an override.
  • Defaults that match 2026 — Empty Latent Image now starts at 1024×1024.
  • Partner / polish: Meshy 7.1, GPT Image 2 transparent backgrounds, MiniMax Music 3 CUDA-graph noise fix, ACE-Step VAE crash fix on non-bf16 GPUs, YuE2 CFG control, Wan peak-VRAM drop with comfy-kitchen attention.

MoGe 3: Depth and Normals That Keep Thin Structure

MoGe (Monocular Geometry) is the tool you reach for when a single photo needs a believable depth map and surface normals — for relighting, camera matching, mesh sketches, or feeding a 3D-aware video pass.

MoGe 3 is the third generation from Microsoft Research (paper: MoGe-3: Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement — track the ComfyUI PR #16381 / CORE-443 for the integration). Instead of decoding a point map only in image space, it lifts a coarse prediction onto a sparse voxel shell and runs a sparse 3D UNet over it. That is the trick that keeps thin structures and sharp depth discontinuities from turning into mush.

In ComfyUI core you get:

  • Official template: utility_moge3_geometry_estimation (one image → depth + normals).
  • Checkpoints under models/geometry_estimation/ — including moge_3_vitg_fp16.safetensors (default) and moge_3_vitl_fp16.safetensors, plus older MoGe 1/2 packs still loadable.
  • New refine_steps widget on Run MoGe Inference / Panorama Inference: default 3, up to 8, or 0 to disable the MoGe-3 refiner (no-op for MoGe 1/2).
  • Panorama path still exists: twelve perspective views → merged depth, plus point-map→mesh and FoV readout for camera match.

Weights are repacked at Comfy-Org/MoGe. The Comfy port reuses FlexGEMM sparse kernels already shipping for Trellis 2 — no exotic extra install on a normal CUDA box.

Creator tip: start with refine_steps=3 on a product or character still. Bump toward 6–8 only when hair, railings, or foliage are collapsing. Zero the refiner when you are batching thumbnails and only need a coarse guide.

Speed and VRAM Quality-of-Life

Two under-the-radar wins will matter more than another shiny sampler:

  1. Qwen text-encoder decode — speculative MTP drafting (widget mtp: auto / off / 2–5) plus graph-friendly KV cache. Same seed will not match non-MTP token-for-token, but the distribution stays correct and walls fall down on prompt rewrite / Text Generate loops.
  2. Fast disk without folklore — if your model drive is a modern NVMe, v0.37 turns on disk-backed dynamic loading for you. Mixed libraries (NVMe + spinning rust) keep RAM priority on the slow side. Force with --fast-disk or kill with --disable-fast-disk when debugging.

Also note: with dynamic VRAM on, the text encoder now stays on GPU instead of commuting back and forth — small change, fewer hitch-spikes mid-graph.

How to Update

# Manager / Desktop users: update to the latest stable.
# Portable / git installs:
cd ComfyUI && git fetch --tags && git checkout v0.37.0
# Then pull MoGe 3 weights into models/geometry_estimation/
# and (if you have not already) Comfy-Org/Qwen-Image-2.1 packs for the image templates.

Open the template browser → search MoGe 3 or Qwen Image 2.1. Desktop and Cloud builds trail the GitHub tag; wait for your channel’s stable bump if you are not on git.

Conclusion

v0.37.0 is a double drop: the Qwen-Image-2.1 story you already know, and a geometry + systems story you should not skip. MoGe 3 gives stills a sharper path into depth and normals; auto fast-disk and faster Qwen encoders make the daily graph less of a waiting room.

Update, drop the MoGe weights, run one photo through the official template, and feel the refiner earn its keep on the thin edges.

—Aurelia ♡