Closed total1ty closed 7 months ago
I just added a third keyframe to the list and now it works a little bit, but ignores the last keyframe. So I added another "dummy" keyframe at the end. The thing is that there is no in-betweens now: With a fixed seed, frame 0-99 are identical and frame 100 changes to the next prompt.
The console log shows the correct strengths and prompts:
G_Clip:
Max Frames: 9999
Current Prompt: beautiful scenery nature glass bottle landscape, , purple galaxy bottle,
Next Prompt: apocalyptic scenery industrial bottle landscape, , green galaxy bottle,
Strength : 0.5800000000000001
L_Clip:
Max Frames: 9999
Current Prompt: beautiful scenery nature glass bottle landscape, , purple galaxy bottle,
Next Prompt: apocalyptic scenery industrial bottle landscape, , green galaxy bottle,
Strength : 0.5800000000000001
100%|███████████████████████████████████████████████████████████████████████████████| 20/20 [00:03<00:00, 6.25it/s]
Prompt executed in 4.15 seconds
got prompt
G_Clip:
Max Frames: 9999
Current Prompt: apocalyptic scenery industrial bottle landscape, , green galaxy bottle,
Next Prompt: arctic scenery bottle landscape, , white galaxy bottle,
Strength : 0.63
L_Clip:
Max Frames: 9999
Current Prompt: apocalyptic scenery industrial bottle landscape, , green galaxy bottle,
Next Prompt: arctic scenery bottle landscape, , white galaxy bottle,
Strength : 0.63
I dont know, maybe this could be the same reason for https://github.com/FizzleDorf/ComfyUI_FizzNodes/issues/73
Here is the updated test setup:
By the way, I noticed that with a fixed seed, Prompt Schedule and Prompt Schedule SDXL deliver pixel-perfectly identical results with the same settings. What's the difference between the nodes except being able to specify separate Text_G and Text_L? (which nobody seems to do anyway :D)
a696099 fixes this issue. you should be able to get a proper result in your examples. thank you for providing them!
Title says it all, the node always acts like current_frame is 0.
I've investigated this using a setup where I can A/B test the regular Prompt Schedule node vs. the SDXL, and the regular one works just fine using exactly the same inputs and prompts. So formatting errors in the scheduling are out of the question.
Frankly, I'm not exactly sure if I'm using the SDXL node correctly. But since changing the prompt at 0 works fine I suspect the node doing (un)fancy things.
Here is my testing setup, just reconnect the Reroutes before the sampler and edit the cyan frame# node between 0/100: