dreamlite-comfyui-lowvram

DreamLite ComfyUI LowVRAM

Run DreamLite-base and DreamLite-mobile from ComfyUI on constrained GPUs, with a practical low-VRAM path for DreamLite-base and tiled RealESRGAN upscaling.

Canonical source: github.com/ENUMERA8OR/dreamlite-comfyui-lowvram

This repository does not host DreamLite or upscaler weights. It contains ComfyUI custom nodes, workflows, documentation, implementation notes, and example outputs.

This project grew out of testing DreamLite on a 4 GB GTX 1650 Ti. The main breakthrough is making DreamLite-base usable at 1024x1024 through:

sequential CPU offload + float32 + GQA-aware query-token attention slicing

Generic Diffusers attention slicing can break DreamLite-base grouped-query attention. The bundled inference path slices the query-token dimension while preserving the grouped-query head layout.

Results

Generated and upscaled examples from the workflow:

More experimental outputs are kept in Generated_Images and raw notes are kept in implementation_notes.

What Is Included

• DreamLite Base Generate V2
• DreamLite Mobile Generate V2
• DreamLite Tiled Upscale
• A ready-to-load ComfyUI workflow:

dreamlite_base_mobile_upscale_compare_workflow_v3_clean.json

The workflow has three lanes:

• DreamLite-base generation -> raw save -> tiled ESRGAN upscale.
• DreamLite-mobile generation -> raw save -> tiled ESRGAN upscale.
• Upload any existing image -> tiled ESRGAN upscale.

Why This Exists

DreamLite-mobile is much easier to run, but DreamLite-base can produce different and often stronger image structure. On 4 GB VRAM, normal DreamLite-base CUDA loading fails. This repo packages the path that worked:

• keep DreamLite-base numerically stable with float32;
• avoid full CUDA residency with sequential_cpu_offload;
• reduce attention peak memory with a DreamLite-compatible GQA query slicer;
• keep upscaling usable with explicit tiled ESRGAN processing.

Repository Layout

comfyui-dreamlite-base-node/
  dreamlite_base_node.py
  scripts/infer_base_offload.py
  examples/

comfyui-dreamlite-mobile-node/
  dreamlite_mobile_node.py
  examples/

comfyui-dreamlite-upscale-node/
  dreamlite_tiled_upscale_node.py

Generated_Images/
  generated and upscaled examples

implementation_notes/
  raw local experiment notes and debugging logs

Requirements

You need an existing local DreamLite checkout and local DreamLite model weights. This repository does not include model weights.

Set these before starting ComfyUI:

export DREAMLITE_PYTHON="/path/to/dreamlite/env/bin/python"
export DREAMLITE_REPO="/path/to/DreamLite"
export DREAMLITE_BASE_MODEL="/path/to/DreamLite-base"
export DREAMLITE_MOBILE_MODEL="/path/to/DreamLite-mobile"

Optional:

export DREAMLITE_TMP="/path/to/writable/tmp"
export DREAMLITE_TORCH_CACHE="/path/to/torch/cache"
export DREAMLITE_OFFLOAD_SCRIPT="/path/to/infer_base_offload.py"
export DREAMLITE_MOBILE_DEVICE="cuda"
export DREAMLITE_MOBILE_MEMORY_MODE="sequential_cpu_offload"

If DREAMLITE_OFFLOAD_SCRIPT is not set, the base node uses:

comfyui-dreamlite-base-node/scripts/infer_base_offload.py

Install

Copy or symlink the node folders into ComfyUI's custom_nodes directory:

ln -s /path/to/this-repo/comfyui-dreamlite-base-node /path/to/ComfyUI/custom_nodes/comfyui-dreamlite-base-node
ln -s /path/to/this-repo/comfyui-dreamlite-mobile-node /path/to/ComfyUI/custom_nodes/comfyui-dreamlite-mobile-node
ln -s /path/to/this-repo/comfyui-dreamlite-upscale-node /path/to/ComfyUI/custom_nodes/comfyui-dreamlite-upscale-node

Restart ComfyUI after setting the environment variables.

Then load:

dreamlite_base_mobile_upscale_compare_workflow_v3_clean.json

Recommended Settings

For a 4 GB GPU, the strongest tested DreamLite-base profile was:

memory_mode: sequential_cpu_offload
dtype: float32
attention_mode: gqa_query_slicing
attention_slice_size: 256
resolution: 1024x1024
steps: 26-30

DreamLite-mobile starting point:

steps: 4
resolution: 1024x1024
dtype: bfloat16

Tiled ESRGAN starting point:

tile_size: 256
overlap: 32
output_device: cpu

Upscaling Existing Images

The workflow includes an independent upload lane:

Load Image -> UpscaleModelLoader -> DreamLite Tiled Upscale -> SaveImage

Use this lane when you want to upscale any existing image without running DreamLite generation first.

Place ESRGAN or RealESRGAN-compatible weights in:

ComfyUI/models/upscale_models

The tested model was:

RealESRGAN_x4plus.pth

Notes

• This repo does not include DreamLite model weights.
• This repo does not include upscaler weights.
• implementation_notes/ intentionally preserves raw experiment notes. Some notes may mention local machine paths; they are not installation instructions.
• DreamLite, ComfyUI, model checkpoints, and upscaler checkpoints are separate projects with their own licenses.

License

Apache-2.0. See LICENSE and NOTICE.