haylan/LLM-Server
1
0
Fork 0
Code Issues 9 Pull Requests Actions Packages Projects Releases Wiki Activity

Best diffusion model given full ~32GB headroom (Qwen stopped) #39

New Issue
Closed
opened 2026-09-05 17:36:34 +00:00 by haylan · 1 comment
haylan commented 2026-09-05 17:36:34 +00:00
Owner
Copy Link
Copy Source

Part of #38

Question

With Qwen/llama-server fully stopped, ComfyUI has the full ~32GB R9700 to
itself instead of the ~6GB left over while Qwen runs. Given that headroom,
which model should this map actually build against — stay on FLUX.1-schnell
(Apache-2.0, GGUF Q4_K_S ~7GB, already confirmed to fit even alongside Qwen),
or move up to FLUX.1-dev, SD3.5-large, Qwen-Image, HunyuanImage-3.0, or
Krea-2 (see docs/research/image-generation-options.md for the prior survey
and licenses)? Resolve license fit (homelab, non-commercial is fine),
ROCm/gfx1201 compatibility, and realistic VRAM (fp16/fp8/GGUF) against the
~32GB ceiling — not the ~6GB one. Output: one recommended model + precision,
citing primary sources.

Part of #38 ## Question With Qwen/llama-server fully stopped, ComfyUI has the full ~32GB R9700 to itself instead of the ~6GB left over while Qwen runs. Given that headroom, which model should this map actually build against — stay on FLUX.1-schnell (Apache-2.0, GGUF Q4_K_S ~7GB, already confirmed to fit even alongside Qwen), or move up to FLUX.1-dev, SD3.5-large, Qwen-Image, HunyuanImage-3.0, or Krea-2 (see docs/research/image-generation-options.md for the prior survey and licenses)? Resolve license fit (homelab, non-commercial is fine), ROCm/gfx1201 compatibility, and realistic VRAM (fp16/fp8/GGUF) against the ~32GB ceiling — not the ~6GB one. Output: one recommended model + precision, citing primary sources.
haylan added the wayfinder:research label 2026-09-05 17:36:34 +00:00
haylan added a new dependency 2026-09-05 18:17:47 +00:00
#42 Extend downloader for the diffusion model weights
haylan commented 2026-09-05 19:16:31 +00:00
Author
Owner
Copy Link
Copy Source

Recommend Qwen-Image at FP8 precision (~25GB: fp8 diffusion weights + fp8 text encoder) as the upgrade from FLUX.1-schnell now that the full ~32GB is available. It has the cleanest license (Apache-2.0, no restrictions) and is the only candidate with a ComfyUI workflow pre-validated specifically on this GPU architecture (gfx1201/R9700, via kyuz0/amd-r9700-comfy). HunyuanImage-3.0 is ruled out (CUDA-only, multi-GPU datacenter VRAM floor); Krea-2 is promising but too new for R9700-specific field evidence yet. Full writeup: docs/research/image-generation-model-choice.md

Recommend **Qwen-Image at FP8 precision** (~25GB: fp8 diffusion weights + fp8 text encoder) as the upgrade from FLUX.1-schnell now that the full ~32GB is available. It has the cleanest license (Apache-2.0, no restrictions) and is the only candidate with a ComfyUI workflow pre-validated specifically on this GPU architecture (gfx1201/R9700, via kyuz0/amd-r9700-comfy). HunyuanImage-3.0 is ruled out (CUDA-only, multi-GPU datacenter VRAM floor); Krea-2 is promising but too new for R9700-specific field evidence yet. Full writeup: docs/research/image-generation-model-choice.md
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
wayfinder:grilling
Wayfinder grilling ticket
 wayfinder:map
Wayfinder map issue
 wayfinder:prototype
Wayfinder prototype ticket
 wayfinder:research
Wayfinder research ticket
 wayfinder:task
Wayfinder task ticket
No labels wayfinder:research
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
haylan (Arthur Erlich)
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Blocks
#42 Extend downloader for the diffusion model weights
haylan/LLM-Server
Reference: haylan/LLM-Server#39