Emu 3.5 Emergency Reset

At some point with Emu 3.5 you will hit a wall: an OOM mid-generation, a workflow that refuses to start because the allocator thinks VRAM is still full, or just the sinking feeling that your 4090 has adopted every orphaned tensor it has ever seen. Emu 3.5 Emergency Reset is the node you drag in when that happens. It's the pack's nuclear-ish option - an aggressive, multi-pass attempt to un-stick the CUDA state without restarting ComfyUI.

The name is not an exaggeration about the mechanism. When you run it, it synchronizes the GPU to finish pending ops, runs five garbage-collection passes, empties the CUDA cache, flushes IPC buffers, resets both peak and accumulated memory stats, then walks gc.get_objects() hunting for CUDA tensors with no strong references - orphans - and reports what it found. Everything it does is logged to a reset_log string and to the console so you can see how much VRAM it clawed back (it tracks allocated vs. reserved before and after).

The inputs that matter

• confirm_reset - a BOOLEAN that defaults to false. You must flip it to true for the node to do anything; otherwise it just returns "Reset cancelled". This is a deliberate safety switch so an accidental rerun doesn't flush your GPU state mid-workflow. Forgetting it is the single most common "why isn't it working" moment with this node.
• target_gpu - optional INT, default 0, range 0–7. Picks which CUDA device to reset if you're on a multi-GPU box.

Output

reset_log (STRING) - a multi-line report of every step, ending with a "Freed: X GB" summary. It's an output node, so mostly you read the log in the console or show the string on canvas rather than wiring it onward.

Install and honest expectations

Same pack, same install as the rest:

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/EricRollei/Emu35-Comfyui-Nodes.git emu35
cd emu35
pip install -r requirements.txt

Weights live in ComfyUI/models/emu35/ - BAAI/Emu3.5-Image (34B, ~65GB BF16) plus the vision tokenizer, or the NF4 quantized wikeeyang/Emu35-Image-NF4 for 24GB cards.

Now the honest part: this node cleans up after a mess, it doesn't prevent one. It won't reload your model and it can't conjure VRAM that a resident 34B model is legitimately using. And note the real-world pattern behind OOMs here: on a 24GB card, people often get through model load and then crash ~30% into inference - the fix that actually worked in the community was ComfyUI's --lowvram launch flag with --reserved-vram tuned upward, plus keeping the vision tokenizer on CPU. Emergency Reset is the recovery tool, not the fix. Use it after a crash to get back to a clean slate, then go apply one of those.