Immediate evaluation of variables does not reset between prompts

[tscott65]

I have the following prompt (and corresponding wildcard files for f_sentient and f_animal):

${animal=!{__f_animal__}} ${sentient=!{__f_sentient__}} [${animal}|${sentient}|${animal}] AND ${sentient} in an ancient forest. Dramatic lighting. dappled lighting. deeps shadows.

Every time I run with a "batch count" > 1, each resulting image uses exactly the same wildcard values. The same is true if I set batch-size > 1 and batch_count = 1. The values seem to repeat no matter what.

However, if I hit "Generate" again, the values are different.

Dynamic Prompts version 45b21373 | Sat Jun 3 10:48:47 2023

version: v1.3.2  •  python: 3.10.7  •  torch: 2.0.1+cu118  •  xformers: N/A  •  gradio: 3.32.0  •  checkpoint: 914077285245b21373 Sat Jun 3 10:48:47 2023

[layluke]

@tscott65 Instead of using batch count and batch size, under dynamic prompts, enable "Combinatorial generation" and leverage "Max generations" and "Combinatorial batches".

You may still have issues, A few of us are having issues using Immediate evaluation of variables like you are: https://github.com/adieyal/sd-dynamic-prompts/issues/421

[amurorei01]

I encountered the same problem after update today. And the "batch count" would automatically be 64 (after rebooting it becomes 16), regardless the number setted. Reinstalling dosen't help.

[adieyal]

@akx what are your thoughts on this? It isn't a bug per se in that the sampler is designed to fix the value of the variable once it has been evaluated but the expectation of randomised values within a batch is reasonable. This problem would be resolved if we created a new sampler for every batch but then other session-based state like cyclical samplers wouldn't work. Perhaps we should reset variables each time we sample?

[AIrtistry]

same bug for me

[akx]

@adieyal Hm, don't we reset (or at least overwrite) them in the context for each generation? 🤔

[adieyal]

I'm not 100% sure. The variable assignment is takes place in

https://github.com/adieyal/dynamicprompts/blob/main/src/dynamicprompts/samplers/base.py#L71-L75

    def _get_sequence(
        self,
        command: SequenceCommand,
        context: SamplingContext,
    ) -> StringGen:
        tokens, context = context.process_variable_assignments(command.tokens)
        sub_generators = [context.generator_from_command(c) for c in tokens]


        while True:
            yield rotate_and_join(sub_generators, separator=command.separator)

the variables are processed outside of the loop on line 71 but the generators are sampled inside of the loop and yielded to the caller. From my reading, I think that variables are only ever processed once.

Here is a unit test that fails

def test_random_variables_across_prompts(wildcard_manager: WildcardManager):
    cmd = parse("${ball=!{red|blue}} ${ball} ${ball} ${ball}")
    scon = SamplingContext(
        default_sampling_method=SamplingMethod.RANDOM,
        wildcard_manager=wildcard_manager
    )
    gen = scon.sample_prompts(cmd)
    seen = set(islice(gen, 40))
    assert len(seen) == 2

This change lets the test pass

    def _get_sequence(
        self,
        command: SequenceCommand,
        context: SamplingContext,
    ) -> StringGen:

        while True:
            tokens, context = context.process_variable_assignments(command.tokens)
            sub_generators = [context.generator_from_command(c) for c in tokens]
            yield rotate_and_join(sub_generators, separator=command.separator)

but breaks a lot of other stuff

[anothertal3]

Not sure if I understand this topic correctly but I'm wondering whether #668 is related?