Files
comfyui-workflows/optimized/T2I/flux/README.md
T

58 lines
3.3 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Flux.2 Klein 9B — optimized T2I
This is the current best tested Flux.2 Klein text-to-image workflow for the 8 GB RTX 3070.
Files:
- `workflow.json` — API-format prompt for `/prompt` automation.
- `ui-workflow.json` — native ComfyUI workflow; load it from the web UI with **Load**.
## Settings
- Model: `flux-2-klein-9b-fp8mixed.safetensors`
- Text encoder: `qwen_3_8b_fp8mixed.safetensors` (`flux2`)
- VAE: `flux2-vae.safetensors`
- Resolution: 1080×1920, 9:16 portrait
- Sampler: `euler_cfg_pp`
- Steps: 6
- CFG: 1.3
- Seed: `20260920`
- Measured generation time: approximately 220.3 seconds for the current best (F-T2I-03).
The included seed and prompt are a controlled baseline example. For reuse, replace the prompt, seed, and resolution; sampler, scheduler, steps, and CFG are the optimization controls.
The positive prompt is the shared T2I comparison prompt in [`../prompt.txt`](../prompt.txt), also used by the Qwen workflow.
## Planned optimization matrix
All rows are intentionally **Untested**. Keep the prompt, seed, and resolution from [`../variables.json`](../variables.json) fixed. Flux2Scheduler does not expose a selectable scheduler, so that value remains fixed.
| ID | CFG | Steps | Sampler | Scheduler | Runtime (s) | Tested? | Current Best |
|---|---:|---:|---|---|---:|---|---|
| F-T2I-01 | 1.3 | 6 | euler_cfg_pp | Flux2Scheduler | 263.9 | Yes | No |
| F-T2I-02 | 1.0 | 6 | euler_cfg_pp | Flux2Scheduler | 252.0 | Yes | No |
| F-T2I-03 | 1.1 | 6 | euler_cfg_pp | Flux2Scheduler | 220.3 | Yes | Yes |
| F-T2I-04 | 1.2 | 6 | euler_cfg_pp | Flux2Scheduler | 239.7 | Yes | No |
| F-T2I-05 | 1.3 | 6 | euler_cfg_pp | Flux2Scheduler | 10.1 | Yes | No |
| F-T2I-06 | 1.4 | 6 | euler_cfg_pp | Flux2Scheduler | 156.2 | Yes | No |
| F-T2I-07 | 1.5 | 6 | euler_cfg_pp | Flux2Scheduler | 180.3 | Yes | No |
| F-T2I-08 | 1.3 | 6 | euler | Flux2Scheduler | 172.3 | Yes | No |
| F-T2I-09 | 1.3 | 6 | euler_ancestral | Flux2Scheduler | 220.3 | Yes | No |
| F-T2I-10 | 1.3 | 6 | euler_ancestral_cfg_pp | Flux2Scheduler | 224.3 | Yes | No |
| F-T2I-11 | 1.3 | 6 | res_multistep | Flux2Scheduler | 180.3 | Yes | No |
| F-T2I-12 | 1.3 | 6 | res_multistep_cfg_pp | Flux2Scheduler | 124.2 | Yes | No |
| F-T2I-13 | 1.3 | 4 | euler_cfg_pp | Flux2Scheduler | 122.2 | Yes | No |
| F-T2I-14 | 1.3 | 5 | euler_cfg_pp | Flux2Scheduler | 178.4 | Yes | No |
| F-T2I-15 | 1.3 | 7 | euler_cfg_pp | Flux2Scheduler | 226.3 | Yes | No |
| F-T2I-16 | 1.3 | 8 | euler_cfg_pp | Flux2Scheduler | 282.4 | Yes | No |
| F-T2I-17 | 1.3 | 4 | euler_ancestral | Flux2Scheduler | 70.4 | Yes | No |
| F-T2I-18 | 1.3 | 4 | res_multistep | Flux2Scheduler | 148.2 | Yes | No |
F-T2I-12 was visually rejected as a corrupted/highly oversaturated output. The remaining completed tests were valid but did not improve on F-T2I-03. F-T2I-05's unusually short runtime is a warm/cache measurement and should not be treated as representative of a cold run.
Evaluation complete: **F-T2I-03 remains the current best**. The closest alternatives were F-T2I-05 and F-T2I-15, but neither improved the overall balance of portrait realism, composition, scoreboard stability, and detail.
## Optional second pass
SeedVR2 3B FP8 produced a successful 2560×1440 output from this baseline. A 3840×2160 attempt ran out of the GPU's 8 GB VRAM. The SeedVR2 pass is therefore an optional export step, not part of the fast generation baseline.