#FromTo & embeds

1 messages · Page 1 of 1 (latest)

thick hedge
#

I cannot seem to get <fromto[x.y]:a,b> to work when a and/or b are <embed> token groups.

E.g., I think the following five should be equivalent, but I do not get a pure application (or any?) of the embed in the second through fifth cases (even though they differ a bit):

Works:
embed:_offload_I/zPDXLrl.pt,
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_furry, professional anthro shark portrait, dramatic lighting, dark background,wearing business suit

Does not seem to work as expected:
<fromto[0]:embed:_offload_I/zPDXLrl.pt,embed:_offload_I/zPDXLrl.pt>,
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_furry, professional anthro shark portrait, dramatic lighting, dark background,wearing business suit

<fromto[0]:embed:_offload_I/zPDXLrl.pt|embed:_offload_I/zPDXLrl.pt>,
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_furry, professional anthro shark portrait, dramatic lighting, dark background,wearing business suit

<fromto[1]:embed:_offload_I/zPDXLrl.pt,embed:_offload_I/zPDXLrl.pt>,
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_furry, professional anthro shark portrait, dramatic lighting, dark background,wearing business suit

<fromto[1]:embed:_offload_I/zPDXLrl.pt|embed:_offload_I/zPDXLrl.pt>,
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_furry, professional anthro shark portrait, dramatic lighting, dark background,wearing business suit

I varied the start and end from 0 to 1 (I also tried values in between) and tried both "," and "|" as the divider between a and b (as per docs), where a and b were both the same embed.

astral ether
#

SwarmClipTextEncodeAdvanced was not very robust for dense merges of multiple syntaxes that have to be parsed in the python apparently

#

completely rewrote the logic to allow infinite layering and support for embeddings

thick hedge
#

Lol. I was watching your posting last night (but notably not on this thread and not in commit) ... and I was imagining you were looking at this issue and grumbling, lol.

#

Thanks. 🙂

thick hedge
#

I see two additional oddities:

  1. If a or b is blank, something still isn't working right (try: <fromto[0.9]:,bear> vs <fromto[0.9]:.,bear> ). My workaround will be to use a "." for a or b if I want it blank...should be close enough.
  2. Seems like b is always shifted rightward n tokens even though a and b shouldn't be coexisting in the same list of tokens (maybe a is being replaced by null tokens instead of being excised?). I notice this when stacking <break> statements, it seems b is always combined/merged later in the prompt text. Maybe I am hallucinating?
#

Grid issue:

  1. This seems broken in grid, the second replace of the weights generates all four of the first replace when 0.0 = 0.0, but then only generates four copies of the first replace for 0.1, 0.2, ... 1.0:
Prompt:
<fromto[0.0]:<embed:_offload_I/zPDXLrl.pt>,<embed:_offload_I/zPDXLrl.pt>>
<break>anthro wearing a business suit

Prompt Replace - Parameter #1:
<fromto[0.0]:<embed:_offload_I/zPDXLrl.pt>,<embed:_offload_I/zPDXLrl.pt>>||<fromto[0.0]:<embed:_offload_I/zPDXLrl.pt>,.>||<fromto[0.0]:.,<embed:_offload_I/zPDXLrl.pt>>||<fromto[0.0]:.,.>

Prompt Replace - Parameter #2:
0.0||0.1||0.2||0.3||0.4||0.5||0.6||0.7||0.8||0.9||1
astral ether
#

2 idek what you think is happening 0.o

#

fromto will generate multiple copies of the prompt with every unique text combination, then encode them and store what time stamps they go to, and <break> is in a lower level of the encoder

#

re the grid bit, I'm... not entirely clear on what you're doing, but offhand guess is just the order, possibly uncheck Allow Reordering even since you're trying to stack replaces and that's order sensitive

thick hedge
#

2: I'll see if I can make something repoducible to discuss.
re 3/grid: see if you can reproduce, other than needing a pony model, info is there. I'll try the uncheck reordering, but because I also see some inconsistency in the prompt text (I get the same prompt text for all four cases at 0.1, then four identical pictures with the same prompt text at 0.2, 0.3 etc, I think there's an issue there). I get different prompt text for all four at 0.0 (that one does rotate through the four prompts). I think I see what you mean about the reordering, though, yeah maybe.

#

For number 3, toggling the reorder didn't have any impact.

However, I then swapped the order of the parameters, putting the weights in before the fromto modifications and that worked. again, toggling the reorder had no impact there either, but that's fine. So, probably workaround is...set the parameters correctly dummy. 🙂

I think I'm stuck in the inverted order mindset.

I probably should have had the fromto replace start the substring after the ] as to not include the weight part and avoided the confusion.

thick hedge
#
  1. In Prompt Replace, generation will not start if the first item in the parameter list, which matches what is in the prompt, is set to SKIP:
astral ether
astral ether
#

the topmost is the innermost loop, lower is wider loop

#

so with your setup there:
it hits 0.0, then does each embed
then does 0.1, then each embed