For anyone getting poor results with the new turbo release
Switch from KSampler to SamplerCustomAdvanced like you would with Flux2. Create a ManualSigmas node and add1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0.0
then drag it to the sigmas input on the scheduler node. use euler
I use a very similar curve
[0] 1.0000
[1] 0.9846
[2] 0.9728
[3] 0.9538
[4] 0.9169
[5] 0.8380
[6] 0.7010
[7] 0.4304
[8] 0.0000
I use a very similar curve
[0] 1.0000
[1] 0.9846
[2] 0.9728
[3] 0.9538
[4] 0.9169
[5] 0.8380
[6] 0.7010
[7] 0.4304
[8] 0.0000
from where? just curious. the one I posted is from https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo/blob/main/model_index.json
custom curve from testing over the years.
night and day difference. before images were muddy and ugly, now they are good. thank you
ig "simple" is the next best available scheduler ?
ig "simple" is the next best available scheduler ?
the workflow templates are updated in comfy with the schedule already set. but mathematically no simple isn't close to the correct sigmas. Not saying you can't get good results with other settings though.
@pwnyourace , could you please post a workflow? It would be helpful for novices. I tried modifying the ComfyUI text-to-image workflow with your instructions, but I don't see how.
KSampler has:
- inputs = model, positive, negative, latent_image
- outputs = LATENT
SamplerCustomAdvanced has:
- inputs = noise, guider, sampler, sigmas, latent_image
- outputs = output and denoised_output.
They have almost nothing in common.
@pwnyourace , could you please post a workflow? It would be helpful for novices. I tried modifying the ComfyUI text-to-image workflow with your instructions, but I don't see how.
KSampler has:
- inputs = model, positive, negative, latent_image
- outputs = LATENT
SamplerCustomAdvanced has:
- inputs = noise, guider, sampler, sigmas, latent_image
- outputs = output and denoised_output.
They have almost nothing in common.
It's possible if you're not on the latest comfyui revision you have old templates. https://github.com/Comfy-Org/workflow_templates/blob/main/templates/image_qwen_image_2_1_turbo_t2i.json
https://raw.githubusercontent.com/Comfy-Org/workflow_templates/refs/heads/main/templates/image_qwen_image_2_1_turbo_t2i.json
right click and save as the raw.github link
you'll have this sampler set up
one small bug in the workflow is that the random_noise seed node isn't connected so make sure you connect it