first item of logits_BxCxV should be uncond_logits_BxCxV

[hopingZ]

No description provided.

[robbiemu]

Per my 30 minute forray into this intriguing 1-liner: This is the standard Classifier-Free Guidance (CFG) formula. You would very likely never want cond_logits_BxCxV when cfg_scale = 0 because the conditional model is usually undertrained (or performs poorly on its own).

Here's some LLM guidance on it:

When cfg_scale = 0 and you use pure conditional logits:

• The model hasn't learned to generate high-quality outputs from the conditional path alone
• It may produce outputs that are too conservative, repetitive, or low-quality
• The conditional model "expects" to be corrected/guided by the CFG process

When cfg_scale = 0 and you use unconditional logits (correct behavior):

• You get outputs from the path that was specifically trained to work independently
• The model generates diverse, reasonable outputs without any text conditioning
• This is the proper "baseline" behavior before applying any guidance

...

cfg_scale = 0 should mean "ignore the text condition entirely". If you use conditional logits when cfg_scale = 0, you're still using the text condition! This breaks the intended meaning of the parameter. The buggy version doesn't give you a way to get pure unconditional generation - even at cfg_scale = 0, you still get conditional outputs. This limits the model's flexibility and breaks user expectations about what the parameter should do.

--

maybe even better would be (For performance):

    if cfg_scale == 0.0:
        # Pure unconditional generation - no guidance
        logits_BxCxV = uncond_logits_BxCxV
    elif cfg_scale == 1.0:
        # Pure conditional generation - no CFG computation needed
        logits_BxCxV = cond_logits_BxCxV
    else:
        # Standard CFG formula: start from unconditional baseline and apply guidance
        logits_BxCxV = uncond_logits_BxCxV + cfg_scale * (cond_logits_BxCxV - uncond_logits_BxCxV)