WASasquatch / was-node-suite-comfyui

An extensive node suite for ComfyUI with over 210 new nodes
MIT License
1.15k stars 170 forks source link

How to increase beams and token limit? #262

Open DarkAlchy opened 11 months ago

DarkAlchy commented 11 months ago

, standing in front of a

That normally means it ran out of tokens but I don't see where to increase it nor a way to get more beams so it is more accurate (temprature, top_p, etc.... too).

DarkAlchy commented 11 months ago

def generate(self, image, sample=False, num_beams=3, max_length=30, min_length=10, top_p=0.9, repetition_penalty=1.0): image_embeds = self.visual_encoder(image)

Do we have to adjust that instead of a front interface?

WASasquatch commented 11 months ago

Yeah I did not expose that, but I get quite longer captions so I feel something else is tripping it up there.

On Fri, Nov 3, 2023, 4:51 PM DarkAlchy @.***> wrote:

def generate(self, image, sample=False, num_beams=3, max_length=30, min_length=10, top_p=0.9, repetition_penalty=1.0): image_embeds = self.visual_encoder(image)

Do we have to adjust that instead of a front interface?

— Reply to this email directly, view it on GitHub https://github.com/WASasquatch/was-node-suite-comfyui/issues/262#issuecomment-1793248513, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAIZEZPMEFCIBEXSDNJOJ7TYCV7QNAVCNFSM6AAAAAA65DQW7CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTOOJTGI2DQNJRGM . You are receiving this because you are subscribed to this thread.Message ID: @.***>

DarkAlchy commented 11 months ago

No matter what I tried it always stopped right there in mid sentence.

WASasquatch commented 11 months ago

No matter what I tried it always stopped right there in mid sentence.

Even increasing beam size? What is the image? Some images just trip out captioning. Like scifi you often get " sci fi fi fi fi fi fi fi fi fi fi fi fi" into infinity until end of limit.

DarkAlchy commented 11 months ago

I have had it do it on some others as well. The problem is this is Blip instead of blip2 or llama which do this far less often.