Implementation:Facebookresearch Audiocraft JASCO set generation params

Overview

JASCO_set_generation_params configures the multi-source classifier-free guidance coefficients and other generation parameters for JASCO's flow matching generation process.

Implements

Principle:Facebookresearch_Audiocraft_JASCO_Generation_Configuration

API Signature

def set_generation_params(
    self,
    cfg_coef_all: float = 5.0,
    cfg_coef_txt: float = 0.0,
    **kwargs
) -> None

Source: audiocraft/models/jasco.py, lines 66-83

Parameters

Parameter	Type	Default	Description
cfg_coef_all	float	5.0	Weight for the multi-source CFG all-conditions term. Higher values increase adherence to all provided conditions (text + temporal)
cfg_coef_txt	float	0.0	Weight for the text-only CFG term. When non-zero, adds an additional guidance term that pushes generation toward the text description specifically
**kwargs	dict	empty	Additional parameters forwarded to FlowMatchingModel.generate(), such as euler, euler_steps, ode_rtol, ode_atol

Return Value

Returns None. Modifies self.generation_params in place.

How It Works

from audiocraft.models import JASCO

model = JASCO.get_pretrained('facebook/jasco-chords-drums-400M')

# Default configuration: strong all-conditions guidance, no text-only term
model.set_generation_params(cfg_coef_all=5.0, cfg_coef_txt=0.0)

# Stronger text adherence
model.set_generation_params(cfg_coef_all=3.0, cfg_coef_txt=2.0)

# Use Euler integration with specific step count
model.set_generation_params(cfg_coef_all=5.0, euler=True, euler_steps=100)

# Adjust ODE solver tolerance for quality vs. speed trade-off
model.set_generation_params(cfg_coef_all=5.0, ode_rtol=1e-4, ode_atol=1e-4)

The implementation:

def set_generation_params(self, cfg_coef_all=5.0, cfg_coef_txt=0.0, **kwargs):
    self.generation_params = {
        'cfg_coef_all': cfg_coef_all,
        'cfg_coef_txt': cfg_coef_txt
    }
    self.generation_params.update(kwargs)

Forward Pass: How Parameters Are Used

The generation_params dictionary is unpacked into FlowMatchingModel.generate() during the generation call chain:

# In JASCO._generate_tokens():
return self.lm.generate(
    prompt_tokens, attributes,
    callback=callback, max_gen_len=total_gen_len,
    **self.generation_params  # cfg_coef_all, cfg_coef_txt, plus any kwargs
)

Inside FlowMatchingModel.generate(), these parameters control the multi-source CFG preprocessing:

# In FlowMatchingModel.generate():
condition_tensors, cfg_terms = self._multi_source_cfg_preprocess(
    conditions, cfg_coef_all, cfg_coef_txt
)

Passthrough Parameters

Additional **kwargs that can be passed to control the ODE solver:

Parameter	Type	Default	Description
euler	bool	False	Use fixed-step Euler integration instead of adaptive ODE solver
euler_steps	int	100	Number of Euler steps when euler=True
ode_rtol	float	1e-5	Relative tolerance for the adaptive ODE solver (dopri5)
ode_atol	float	1e-5	Absolute tolerance for the adaptive ODE solver (dopri5)

Comparison with MusicGen Configuration

Aspect	MusicGen	JASCO
CFG coefficients	Single cfg_coef	Multi-source: cfg_coef_all + cfg_coef_txt
Sampling params	use_sampling, top_k, top_p, temperature	Not applicable (continuous, not discrete)
Duration control	User-configurable duration	Fixed from training config (segment_duration)
Generation method	Autoregressive token sampling	ODE integration (Euler or adaptive)

Related Implementations

• Implementation:Facebookresearch_Audiocraft_JASCO_get_pretrained
• Implementation:Facebookresearch_Audiocraft_JASCO_generate_music
• Heuristic:Facebookresearch_Audiocraft_Generation_Parameter_Defaults