ComfyUI prompt control

Nodes for LoRA and prompt scheduling that make basic operations in ComfyUI completely prompt-controllable.

LoRA and prompt scheduling should produce identical output to the equivalent ComfyUI workflow using multiple samplers or the various conditioning manipulation nodes. If you find situations where this is not the case, please report a bug.

What can it do?

Things you can control via the prompt:

• Prompt editing and filtering without multiple samplers
• LoRA loading and scheduling (including LoRA block weights)
• Prompt masking and area control, combining prompts and interpolation
• SDXL parameters
• Other miscellaneous things

This example workflow implements a two-pass workflow illustrating most scheduling features.

The tools in this repository combine well with the macro and wildcard functionality in comfyui-utility-nodes

Requirements

You need to have lark installed in your Python environment for parsing to work (If you reuse A1111's venv, it'll already be there)

If you use the portable version of ComfyUI on Windows with its embedded Python, you must open a terminal in the ComfyUI installation directory and run the command:

.\python_embeded\python.exe -m pip install lark

Then restart ComfyUI afterwards.

Notable changes

I try to avoid behavioural changes that break old prompts, but they may happen occasionally.

• 2024-02-02 The node will now automatically enable offloading LoRA backup weights to the CPU if you run out of memory during LoRA operations, even when --highvram is specified. This change persists until ComfyUI is restarted.
• 2024-01-14 Multiple CLIP_L instances are now joined with a space separator instead of concatenated.
• 2024-01-09 AITemplate support dropped. I don't recommend or test AITemplate anymore. Use Stable-Fast instead (see below for info)
• 2024-01-08 Prompt control now enables in-place weight updates on the model. This shouldn't affect anything, but increases performance slightly. You can disable this by setting the environment variable PC_NO_INPLACE_UPDATE to any non-empty value.
• 2023-12-28 MASK now uses ComfyUI's mask_strength attribute instead of calculating it on its own. This changes its behaviour slightly.
• 2023-12-06: Removed JinjaRender, SimpleWildcard, ConditioningCutoff, CondLinearInterpolate and StringConcat. For the first two, see this repository for mostly-compatible implementations.
• 2023-10-04: STYLE:... syntax changed to STYLE(...)

Note on how schedules work

ComfyUI does not use the step number to determine whether to apply conds; instead, it uses the sampler's timestep value which is affected by the scheduler you're using. This means that when the sampler scheduler isn't linear, the schedules generated by prompt control will not be either.

Currently there doesn't seem to be a good way to change this.

You can try using the PCSplitSampling node to enable an alternative method of sampling.

Scheduling syntax

Syntax is like A1111 for now, but only fractions are supported for steps.

a [large::0.1] [cat|dog:0.05] [<lora:somelora:0.5:0.6>::0.5]
[in a park:in space:0.4]

You can also use a [b:c:0.3,0.7] as a shortcut. The prompt be a until 0.3, a b until 0.7, and then a c. [a:0.1,0.4] is equivalent to [a::0.1,0.4]

LoRA loading

LoRAs can be loaded by referring to the filename without extension and subdirectories will also be searched. For example, <lora:cats:1>. will match both cats.safetensors and sd15/animals/cats.safetensors. If there are multiple LoRAs with the same name, the first match will be loaded.

Alternatively, the name can include the full directory path relative to ComfyUI's search paths, without extension: <lora:XL/sdxllora:0.5>. In this case, the full path must match.

If no match is found, the node will try to replace spaces with underscores and search again. That is, <lora:cats and dogs:1> will find cats_and_dogs.safetensors. This helps with some autocompletion scripts that replace underscores with spaces.

Finally, you can give the exact path (including the extension) as shown in LoRALoader.

Alternating

Alternating syntax is [a|b:pct_steps], causing the prompt to alternate every pct_steps. pct_steps defaults to 0.1 if not specified. You can also have more than two options.

Sequences

The syntax [SEQ:a:N1:b:N2:c:N3] is shorthand for [a:[b:[c::N3]:N2]:N1] ie. it switches from a to b to c to nothing at the specified points in sequence.

Might be useful with Jinja templating (see https://github.com/asagi4/comfyui-utility-nodes). For example:

[SEQ<% for x in steps(0.1, 0.9, 0.1) %>:<lora:test:<= sin(x*pi) + 0.1 =>>:<= x =><% endfor %>]

generates a LoRA schedule based on a sinewave

Tag selection

Using the FilterSchedule node, in addition to step percentages, you can use a tag to select part of an input:

a large [dog:cat<lora:catlora:0.5>:SECOND_PASS]

Set the tags parameter in the FilterSchedule node to filter the prompt. If the tag matches any tag tags (comma-separated), the second option is returned (cat, in this case, with the LoRA). Otherwise, the first option is chosen (dog, without LoRA).

the values in tags are case-insensitive, but the tags in the input must be uppercase A-Z and underscores only, or they won't be recognized. That is, [dog:cat:hr] will not work.

For example, a prompt

a [black:blue:X] [cat:dog:Y] [walking:running:Z] in space

with tags x,z would result in the prompt a blue cat running in space

Prompt interpolation

a red [INT:dog:cat:0.2,0.8:0.05] will attempt to interpolate the tensors for a red dog and a red cat between the specified range in as many steps of 0.05 as will fit.

SDXL

The nodes do not treat SDXL models specially, but there are some utilities that enable SDXL specific functionality.

You can use the function SDXL(width height, target_width target_height, crop_w crop_h) to set SDXL prompt parameters. SDXL() is equivalent to SDXL(1024 1024, 1024 1024, 0 0) unless the default values have been overridden by PCScheduleSettings.

To set the clip_l prompt, as with CLIPTextEncodeSDXL, use the function CLIP_L(prompt text goes here).

Things to note:

• Multiple instances of CLIP_L are joined with a space. That is, CLIP_L(foo)CLIP_L(bar) is the same as CLIP_L(foo bar)
• Using BREAK isn't supported in it; it'll just parse as the plain word BREAK.
• similarly, AND inside CLIP_L does not do anything sensible; CLIP_L(foo AND bar) will parse as two prompts CLIP_L(foo and bar)
• CLIP_L and SDXL have no effect on SD 1.5.
• The rest of the prompt becomes the clip_g prompt.
• If there is no CLIP_L or SDXL, the prompts will work as with CLIPTextEncode.

Other syntax:

• <emb:xyz> is alternative syntax for embedding:xyz to work around a syntax conflict with [embedding:xyz:0.5] which is parsed as a schedule that switches from embedding to xyz.

• The keyword BREAK causes the prompt to be tokenized in separate chunks, which results in each chunk being individually padded to the text encoder's maximum token length. This is mostly equivalent to the ConditioningConcat node.

Combining prompts

AND can be used to combine prompts. You can also use a weight at the end. It does a weighted sum of each prompt,

cat :1 AND dog :2

The weight defaults to 1 and are normalized so that a:2 AND b:2 is equal to a AND b. AND is processed after schedule parsing, so you can change the weight mid-prompt: cat:[1:2:0.5] AND dog

if there is COMFYAND() in the prompt, the behaviour of AND will change to work like ConditioningCombine, but in practice this seems to be just slower while producing the same output.

Functions

There are some "functions" that can be included in a prompt to do various things.

Functions have the form FUNCNAME(param1, param2, ...). How parameters are interpreted is up to the function. Note: Whitespace is not stripped from string parameters by default. Commas can be escaped with \,

Like AND, these functions are parsed after regular scheduling syntax has been expanded, allowing things like [AREA:MASK:0.3](...), in case that's somehow useful.

SHUFFLE and SHIFT

Default parameters: SHUFFLE(seed=0, separator=,, joiner=,), SHIFT(steps=0, separator=,, joiner=,)

SHIFT moves elements to the left by steps. The default is 0 so SHIFT() does nothing SHUFFLE generates a random permutation with seed as its seed.

These functions are applied to each prompt chunk after BREAK, AND etc. have been parsed. The prompt is split by separator, the operation is applied, and it's then joined back by joiner.

Multiple instances of these functions are applied in the order they appear in the prompt.

NOTE: These functions are not smart about syntax and will break emphasis if the separator occurs inside parentheses. I might fix this at some point, but for now, keep this in mind.

For example:

• SHIFT(1) cat, dog, tiger, mouse does a shift and results in dog, tiger, mouse, cat. (whitespace may vary)

• SHIFT(1,;) cat, dog ; tiger, mouse results in tiger, mouse, cat, dog

• SHUFFLE() cat, dog, tiger, mouse results in cat, dog, mouse, tiger

• SHUFFLE() SHIFT(1) cat, dog, tiger, mouse results in dog, mouse, tiger, cat

• SHIFT(1) cat,dog BREAK tiger,mouse results in dog,cat BREAK tiger,mouse

• SHIFT(1) cat, dog AND SHIFT(1) tiger, mouse results in dog, cat BREAK mouse, tiger

Whitespace is not stripped and may also be used as a joiner or separator

• SHIFT(1,, ) cat,dog results in dog cat

NOISE

The function NOISE(weight, seed) adds some random noise into the prompt. The seed is optional, and if not specified, the global RNG is used. weight should be between 0 and 1.

MASK, IMASK and AREA

You can use MASK(x1 x2, y1 y2, weight, op) to specify a region mask for a prompt. The values are specified as a percentage with a float between 0 and 1, or as absolute pixel values (these can't be mixed). 1 will be interpreted as a percentage instead of a pixel value.

Similarly, you can use AREA(x1 x2, y1 y2, weight) to specify an area for the prompt (see ComfyUI's area composition examples). The area is calculated by ComfyUI relative to your latent size.

Custom masks: IMASK and PCScheduleAddMasks

You can attach custom masks to a PROMPT_SCHEDULE with the PCScheduleAddMasks node and then refer to those masks in the prompt using IMASK(index, weight, op). Indexing starts from zero, so 0 is the first attached mask etc. PCSCheduleAddMasks ignores empty inputs, so if you only add a mask to the mask4 input, it will still have index 0.

Applying PCScheduleAddMasks multiple times appends masks to a schedule rather than overriding existing ones, so if you need more than 4, you can just use it more than once.

Behaviour of masks

If multiple MASKs are specified, they are combined together with ComfyUI's MaskComposite node, with op specifying the operation to use (default multiply). In this case, the combined mask weight can be set with MASKW(weight) (defaults to 1.0).

Masks assume a size of (512, 512), unless overridden with PCScheduleSettings and pixel values will be relative to that. ComfyUI will scale the mask to match the image resolution. You can change it manually by using MASK_SIZE(width, height) anywhere in the prompt,

These are handled per AND-ed prompt, so in prompt1 AND MASK(...) prompt2, the mask will only affect prompt2.

The default values are MASK(0 1, 0 1, 1) and you can omit unnecessary ones, that is, MASK(0 0.5, 0.3) is MASK(0 0.5, 0.3 1, 1)

Note that because the default values are percentages, MASK(0 256, 64 512) is valid, but MASK(0 200) will raise an error.

Masking does not affect LoRA scheduling unless you set unet weights to 0 for a LoRA.

FEATHER

When you use MASK or IMASK, you can also call FEATHER(left top right bottom) to apply feathering using ComfyUI's FeatherMask node. The values are in pixels and default to 0.

If multiple masks are used, FEATHER is applied before compositing in the order they appear in the prompt, and any leftovers are applied to the combined mask. If you want to skip feathering a mask while compositing, just use FEATHER() with no arguments.

For example:

MASK(1) MASK(2) MASK(3) FEATHER(1) FEATHER() FEATHER(3) weirdmask FEATHER(4)

gives you a mask that is a combination of 1, 2 and 3, where 1 and 3 are feathered before compositing and then FEATHER(4) is applied to the composite.

The order of the FEATHER and MASK calls doesn't matter; you can have FEATHER before MASK or even interleave them.

Schedulable LoRAs

The ScheduleToModel node patches a model so that when sampling, it'll switch LoRAs between steps. You can apply the LoRA's effect separately to CLIP conditioning and the unet (model).

Swapping LoRAs often can be quite slow without the --highvram switch because ComfyUI will shuffle things between the CPU and GPU. When things stay on the GPU, it's quite fast.

If you run out of VRAM during a LoRA swap, the node will attempt to save VRAM by enabling CPU offloading for future generations even in highvram mode. This persists until ComfyUI is restarted.

You can also set the PC_RETRY_ON_OOM environment variable to any non-empty value to automatically retry sampling once if VRAM runs out.

LoRA Block Weight

If you have ComfyUI Inspire Pack installed, you can use its Lora Block Weight syntax, for example:

a prompt <lora:cars:1:LBW=SD-OUTALL;A=1.0;B=0.0;>

The ; is optional if there is only 1 parameter. The syntax is the same as in the ImpactWildcard node, documented here

Other integrations

Advanced CLIP encoding

You can use the syntax STYLE(weight_interpretation, normalization) in a prompt to affect how prompts are interpreted.

Without any extra nodes, only perp is available, which does the same as ComfyUI_PerpWeight extension.

If you have Advanced CLIP Encoding nodes cloned into your custom_nodes, more options will be available.

The style can be specified separately for each AND:ed prompt, but the first prompt is special; later prompts will "inherit" it as default. For example:

STYLE(A1111) a (red:1.1) cat with (brown:0.9) spots and a long tail AND an (old:0.5) dog AND a (green:1.4) (balloon:1.1)

will interpret everything as A1111, but

a (red:1.1) cat with (brown:0.9) spots and a long tail AND STYLE(A1111) an (old:0.5) dog AND a (green:1.4) (balloon:1.1)

Will interpret the first one using the default ComfyUI behaviour, the second prompt with A1111 and the last prompt with the default again

For things (ie. the code imports) to work, the nodes must be cloned in a directory named exactly ComfyUI_ADV_CLIP_emb.

Cutoff node integration

If you have ComfyUI Cutoff cloned into your custom_nodes, you can use the CUT keyword to use cutoff functionality

The syntax is

a group of animals, [CUT:white cat:white], [CUT:brown dog:brown:0.5:1.0:1.0:_]

the parameters in the CUT section are region_text:target_text:weight;strict_mask:start_from_masked:padding_token of which only the first two are required. If strict_mask, start_from_masked or padding_token are specified in more than one section, the last one takes effect for the whole prompt

Stable-Fast

The prompt control node works well with ComfyUI_stable_fast. However, you should apply ScheduleToModel after applying Apply StableFast Unet to prevent constant recompilations.

Nodes

PromptToSchedule

Parses a schedule from a text prompt. A schedule is essentially an array of (valid_until, prompt) pairs that the other nodes can use.

FilterSchedule

Filters a schedule according to its parameters, removing any changes that do not occur within [start, end).

The node also does tag filtering if any tags are specified.

Always returns at least the last prompt in the schedule if everything would otherwise be filtered.

start=0, end=0 returns the prompt at the start and start=1.0, end=1.0 returns the prompt at the end.

ScheduleToCond

Produces a combined conditioning for the appropriate timesteps. From a schedule. Also applies LoRAs to the CLIP model according to the schedule.

ScheduleToModel

Produces a model that'll cause the sampler to reapply LoRAs at specific steps according to the schedule.

This depends on a callback handled by a monkeypatch of the ComfyUI sampler function, so it might not work with custom samplers, but it shouldn't interfere with them either.

PCSplitSampling

Causes sampling to be split into multiple sampler calls instead of relying on timesteps for scheduling. This makes the schedules more accurate, but seems to cause weird behaviour with SDE samplers. (Upstream bug?)

PCScheduleSettings

Returns an object representing default values for the SDXL function and allows configuring MASK_SIZE outside the prompt. You need to apply them to a schedule with PCApplySettings. Note that for the SDXL settings to apply, you still need to have SDXL() in the prompt.

The "steps" parameter currently does nothing; it's for future features.

PCApplySettings

Applies the give default values from PCScheduleSettings to a schedule

PCPromptFromSchedule

Extracts a text prompt from a schedule; also logs it to the console. LoRAs are not included in the text prompt, though they are logged.

PCScheduleAddMasks

Attaches custom masks to a PROMPT_SCHEDULE that can then be used in a prompt.

PromptControlSimple

This node exists purely for convenience. It's a combination of PromptToSchedule, ScheduleToCond, ScheduleToModel and FilterSchedule such that it provides as output a model, positive conds and negative conds, both with and without any specified filters applied.

This makes it handy for quick one- or two-pass workflows.

Older nodes

• EditableCLIPEncode: A combination of PromptToSchedule and ScheduleToCond
• LoRAScheduler: A combination of PromptToSchedule, FilterSchedule and ScheduleToModel

Known issues

• If you use LoRA scheduling in a workflow with LoRALoader nodes, you might get inconsistent results. For now, just avoid mixing ScheduleToModel or LoRAScheduler with LoRALoader. See https://github.com/asagi4/comfyui-prompt-control/issues/36
• Workflows using SamplerCustom will calculate LoRA schedules based on the number of sigmas given to the sampler instead of the number of steps, since that information isn't available.
• CUT does not work with STYLE:perp
• PCSplitSampling overrides ComfyUI's BrownianTreeNoiseSampler noise sampling behaviour so that each split segment doesn't add crazy amounts of noise to the result with some samplers.
• Split sampling may have weird behaviour if your step percentages go below 1 step.
• Interpolation is probably buggy and will likely change behaviour whenever code gets refactored.
• If execution is interrupted and LoRA scheduling is used, your models might be left in an undefined state until you restart ComfyUI