diff --git a/modules/framepack/framepack_load.py b/modules/framepack/framepack_load.py
index 74e853a52..5b3f09769 100644
--- a/modules/framepack/framepack_load.py
+++ b/modules/framepack/framepack_load.py
@@ -196,7 +196,7 @@ def load_model(variant: str | None = None, pipeline: str | None = None, text_enc
 
         diffusers.loaders.peft._SET_ADAPTER_SCALE_FN_MAPPING['HunyuanVideoTransformer3DModelPacked'] = lambda model_cls, weights: weights # pylint: disable=protected-access
         log.info(f'FramePack load: model={shared.sd_model.__class__.__name__} variant="{variant}" type={shared.sd_model_type} time={t1-t0:.2f}')
-        sd_models.apply_balanced_offload(shared.sd_model)
+        sd_models.set_diffuser_offload(shared.sd_model)
         devices.torch_gc(force=True, reason='load')
 
     except Exception as e:
diff --git a/modules/framepack/framepack_worker.py b/modules/framepack/framepack_worker.py
index 5a0388336..29fb1006b 100644
--- a/modules/framepack/framepack_worker.py
+++ b/modules/framepack/framepack_worker.py
@@ -130,6 +130,7 @@ def worker(
         else:
             end_latent = None
         sd_models.apply_balanced_offload(shared.sd_model)
+        sd_models.offload_ondemand(shared.sd_model, reason='vae encode') # group offload returns the vae through its on-demand placement rather than the balanced seam
         timer.process.add('encode', time.time()-t0)
         shared.state.end(jobid)
         return start_latent, end_latent
@@ -317,6 +318,7 @@ def worker(
                         current_pixels = framepack_vae.vae_decode(real_history_latents[:, :, :section_latent_frames], vae_type=vae_type).cpu()
                         history_pixels = utils.soft_append_bcthw(current_pixels, history_pixels, overlapped_frames)
                 sd_models.apply_balanced_offload(shared.sd_model)
+                sd_models.offload_ondemand(shared.sd_model, reason='vae decode')
                 timer.process.add('vae', time.time()-t_vae)
 
                 if is_last_section:
@@ -376,6 +378,7 @@ def worker(
         errors.display(e, 'FramePack')
 
     sd_models.apply_balanced_offload(shared.sd_model)
+    sd_models.offload_ondemand(shared.sd_model, reason='finish')
     stream.output_queue.push(('end', None))
     t1 = time.time()
     log.info(f'Processed: frames={total_generated_frames} fps={total_generated_frames/(t1-t0):.2f} its={(shared.state.sampling_step)/(t1-t0):.3f} time={t1-t0:.2f} timers={timer.process.dct()} memory={memstats.memory_stats()}')