diff --git a/filters/f_auto_filters.c b/filters/f_auto_filters.c
index 6040d6be3add1..11303e385d63c 100644
--- a/filters/f_auto_filters.c
+++ b/filters/f_auto_filters.c
@@ -63,12 +63,15 @@ static void deint_process(struct mp_filter *f)
     bool filter_needed = opts->deinterlace == 1 ||
                          (opts->deinterlace == -1 && (p->interlaced_frame || p->sub.filter));
 
-    // If the image format changed, or if we no longer need a filter,
-    // destroy any existing filter.
-    if (img->imgfmt != p->prev_imgfmt || (p->sub.filter && !filter_needed)) {
+    // If the image format changed, destroy any existing filter immediately since
+    // it may not support the new format. If we no longer need a filter, drain
+    // and destroy it gracefully.
+    if (img->imgfmt != p->prev_imgfmt) {
+        mp_subfilter_destroy(&p->sub);
+        p->prev_imgfmt = img->imgfmt;
+    } else if (p->sub.filter && !filter_needed) {
         if (!mp_subfilter_drain_destroy(&p->sub))
             return;
-        p->prev_imgfmt = img->imgfmt;
     }
 
     // If no filter is needed or if the filter is already inserted and we reach
@@ -92,6 +95,7 @@ static void deint_process(struct mp_filter *f)
         field_parity = "auto";
     }
 
+    struct mp_stream_info *info = mp_filter_find_stream_info(f);
     bool has_filter = true;
     if (img->imgfmt == IMGFMT_VDPAU) {
         char *args[] = {"deint", "yes",
@@ -123,6 +127,11 @@ static void deint_process(struct mp_filter *f)
                         "parity", field_parity, NULL};
         p->sub.filter =
             mp_create_user_filter(f, MP_OUTPUT_CHAIN_VIDEO, "vavpp", args);
+    } else if (info && info->deinterlace && !IMGFMT_IS_HWACCEL(img->imgfmt)) {
+        char *args[] = {"interlaced-only", opts->deinterlace == 1 ? "no" : "yes",
+                        "parity", field_parity, NULL};
+        p->sub.filter = mp_create_user_filter(f, MP_OUTPUT_CHAIN_VIDEO,
+                                              "fieldrate", args);
     } else {
         has_filter = false;
     }
diff --git a/filters/f_output_chain.c b/filters/f_output_chain.c
index 8ded23177da84..7c9ab99acf55c 100644
--- a/filters/f_output_chain.c
+++ b/filters/f_output_chain.c
@@ -387,6 +387,7 @@ void mp_output_chain_set_vo(struct mp_output_chain *c, struct vo *vo)
     p->stream_info.osd = vo ? vo->osd : NULL;
     p->stream_info.vflip = vo ? vo->driver->caps & VO_CAP_VFLIP : false;
     p->stream_info.rotate90 = vo ? vo->driver->caps & VO_CAP_ROTATE90 : false;
+    p->stream_info.deinterlace = vo ? vo->driver->caps & VO_CAP_DEINTERLACE : false;
     p->stream_info.dr_vo = vo;
     p->vo = vo;
     update_output_caps(p);
diff --git a/filters/filter.h b/filters/filter.h
index 5bf1e48a84958..334c46e64a024 100644
--- a/filters/filter.h
+++ b/filters/filter.h
@@ -408,6 +408,7 @@ struct mp_stream_info {
     bool vflip;
     bool rotate90;
     bool force_swdec;
+    bool deinterlace;
     struct vo *dr_vo; // for calling vo_get_image()
 };
 
diff --git a/filters/user_filters.c b/filters/user_filters.c
index 7bdda6965369f..49b674111f6ff 100644
--- a/filters/user_filters.c
+++ b/filters/user_filters.c
@@ -106,6 +106,7 @@ const struct mp_user_filter_entry *vf_list[] = {
 #if (HAVE_GL && HAVE_EGL) || HAVE_VULKAN
     &vf_gpu,
 #endif
+    &vf_fieldrate,
 };
 
 static bool get_vf_desc(struct m_obj_desc *dst, int index)
diff --git a/filters/user_filters.h b/filters/user_filters.h
index aa88649b33dec..8698ba2c1a099 100644
--- a/filters/user_filters.h
+++ b/filters/user_filters.h
@@ -38,3 +38,4 @@ extern const struct mp_user_filter_entry vf_d3d11vpp;
 extern const struct mp_user_filter_entry vf_amf_frc;
 extern const struct mp_user_filter_entry vf_fingerprint;
 extern const struct mp_user_filter_entry vf_gpu;
+extern const struct mp_user_filter_entry vf_fieldrate;
diff --git a/meson.build b/meson.build
index 1cd2570f0d7ef..1cd90648e64a7 100644
--- a/meson.build
+++ b/meson.build
@@ -217,6 +217,7 @@ sources = files(
     'video/filter/refqueue.c',
     'video/filter/vf_format.c',
     'video/filter/vf_sub.c',
+    'video/filter/vf_fieldrate.c',
     'video/fmt-conversion.c',
     'video/hwdec.c',
     'video/image_loader.c',
diff --git a/video/filter/vf_fieldrate.c b/video/filter/vf_fieldrate.c
new file mode 100644
index 0000000000000..d2d66b30d5c54
--- /dev/null
+++ b/video/filter/vf_fieldrate.c
@@ -0,0 +1,107 @@
+#include "filters/filter_internal.h"
+#include "filters/user_filters.h"
+#include "refqueue.h"
+
+struct opts {
+    int field_parity;
+    bool interlaced_only;
+};
+
+struct priv {
+    struct opts *opts;
+    struct mp_refqueue *queue;
+};
+
+static void vf_field_process(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+    mp_refqueue_execute_reinit(p->queue);
+
+    if (!mp_refqueue_can_output(p->queue))
+        return;
+
+    struct mp_image *in = mp_refqueue_get(p->queue, 0);
+    struct mp_image *out = mp_image_new_ref(in);
+    if (!out) {
+        mp_refqueue_write_out_pin(p->queue, NULL);
+        return;
+    }
+    // This filter does not deinterlace. It only emits one output per field so
+    // the VO gets called at fieldrate cadence.
+    if (mp_refqueue_should_deint(p->queue)) {
+        out->field_tick = mp_refqueue_is_second_field(p->queue) ?
+                          MP_FIELD_TICK_SECOND : MP_FIELD_TICK_FIRST;
+        out->fields |= MP_IMGFIELD_INTERLACED;
+        if (mp_refqueue_top_field_first(p->queue)) {
+            out->fields |= MP_IMGFIELD_TOP_FIRST;
+        } else {
+            out->fields &= ~MP_IMGFIELD_TOP_FIRST;
+        }
+    } else {
+        out->field_tick = MP_FIELD_TICK_NONE;
+    }
+    mp_refqueue_write_out_pin(p->queue, out);
+}
+
+static void vf_field_reset(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+    mp_refqueue_flush(p->queue);
+}
+
+static void vf_field_destroy(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+    mp_refqueue_flush(p->queue);
+    talloc_free(p->queue);
+}
+
+static const struct mp_filter_info vf_field_filter = {
+    .name = "fieldrate",
+    .process = vf_field_process,
+    .reset = vf_field_reset,
+    .destroy = vf_field_destroy,
+    .priv_size = sizeof(struct priv),
+};
+
+static struct mp_filter *vf_field_create(struct mp_filter *parent, void *options)
+{
+    struct mp_filter *f = mp_filter_create(parent, &vf_field_filter);
+    if (!f)
+        return NULL;
+    struct priv *p = f->priv;
+    p->opts = talloc_steal(p, options);
+
+    mp_filter_add_pin(f, MP_PIN_IN, "in");
+    mp_filter_add_pin(f, MP_PIN_OUT, "out");
+
+    p->queue = mp_refqueue_alloc(f);
+
+    mp_refqueue_set_refs(p->queue, 0, 0);
+    mp_refqueue_set_mode(p->queue,
+        MP_MODE_DEINT |
+        MP_MODE_OUTPUT_FIELDS |
+        (p->opts->interlaced_only ? MP_MODE_INTERLACED_ONLY : 0));
+    mp_refqueue_set_parity(p->queue, p->opts->field_parity);
+
+    return f;
+}
+#define OPT_BASE_STRUCT struct opts
+static const m_option_t vf_opts_fields[] = {
+    {"interlaced-only", OPT_BOOL(interlaced_only)},
+    {"parity", OPT_CHOICE(field_parity,
+        {"tff", MP_FIELD_PARITY_TFF},
+        {"bff", MP_FIELD_PARITY_BFF},
+        {"auto", MP_FIELD_PARITY_AUTO})},
+    {0}
+};
+
+const struct mp_user_filter_entry vf_fieldrate = {
+    .desc = {
+        .description = "Emit one frame per field",
+        .name = "fieldrate",
+        .priv_size = sizeof(OPT_BASE_STRUCT),
+        .options = vf_opts_fields,
+    },
+    .create = vf_field_create,
+};
diff --git a/video/mp_image.h b/video/mp_image.h
index 5fe523dd64e96..6ad78091e39a5 100644
--- a/video/mp_image.h
+++ b/video/mp_image.h
@@ -38,6 +38,12 @@
 #define MP_IMGFIELD_REPEAT_FIRST 0x04
 #define MP_IMGFIELD_INTERLACED 0x20
 
+enum mp_image_field_tick {
+    MP_FIELD_TICK_NONE = 0,
+    MP_FIELD_TICK_FIRST,
+    MP_FIELD_TICK_SECOND,
+};
+
 // Describes image parameters that usually stay constant.
 // New fields can be added in the future. Code changing the parameters should
 // usually copy the whole struct, so that fields added later will be preserved.
@@ -97,6 +103,8 @@ typedef struct mp_image {
     int pict_type; // 0->unknown, 1->I, 2->P, 3->B
     int fields;
 
+    enum mp_image_field_tick field_tick;
+
     /* only inside filter chain */
     double pts;
     /* only after decoder */
diff --git a/video/out/vo.h b/video/out/vo.h
index a98b0f9496853..89afa9e6731b9 100644
--- a/video/out/vo.h
+++ b/video/out/vo.h
@@ -205,6 +205,8 @@ enum {
     VO_CAP_FRAMEOWNER   = 1 << 5,
     // VO does handle mp_image_params.vflip
     VO_CAP_VFLIP        = 1 << 6,
+    // VO supports deinterlacing
+    VO_CAP_DEINTERLACE  = 1 << 7,
 };
 
 enum {
diff --git a/video/out/vo_gpu_next.c b/video/out/vo_gpu_next.c
index 308a9fb02f5b6..3731663c34adc 100644
--- a/video/out/vo_gpu_next.c
+++ b/video/out/vo_gpu_next.c
@@ -147,6 +147,7 @@ struct priv {
     pl_options pars;
     struct m_config_cache *opts_cache;
     struct m_config_cache *next_opts_cache;
+    struct m_config_cache *filter_opts_cache;
     struct gl_next_opts *next_opts;
     struct cache shader_cache, icc_cache;
     struct mp_csp_equalizer_state *video_eq;
@@ -890,6 +891,7 @@ static void update_options(struct vo *vo)
     pl_options pars = p->pars;
     bool changed = m_config_cache_update(p->opts_cache);
     changed = m_config_cache_update(p->next_opts_cache) || changed;
+    changed = m_config_cache_update(p->filter_opts_cache) || changed;
     if (changed)
         update_render_options(vo);
 
@@ -1143,13 +1145,29 @@ static bool draw_frame(struct vo *vo, struct vo_frame *frame)
         mpi->priv = fp;
         fp->vo = vo;
 
+        bool use_fields = params.deinterlace_params &&
+                          mpi->fields & MP_IMGFIELD_INTERLACED &&
+                          !IMGFMT_IS_HWACCEL(mpi->imgfmt);
+        // vf_fieldrate emits a frame for each field only to make mpv render
+        // at the second field PTS. But don't actually push it to pl_queue,
+        // because it already derives it internally from first_field.
+        if (use_fields && mpi->field_tick == MP_FIELD_TICK_SECOND) {
+            talloc_free(mpi);
+            p->last_id = id;
+            continue;
+        }
+        int first_field = !use_fields ? PL_FIELD_NONE :
+            (mpi->fields & MP_IMGFIELD_TOP_FIRST) ? PL_FIELD_TOP : PL_FIELD_BOTTOM;
+
         pl_queue_push(p->queue, &(struct pl_source_frame) {
             .pts = mpi->pts,
-            .duration = can_interpolate ? frame->approx_duration : 0,
+            .duration = can_interpolate ? frame->approx_duration :
+                        use_fields ? mpi->pkt_duration : 0,
             .frame_data = mpi,
             .map = map_frame,
             .unmap = unmap_frame,
             .discard = discard_frame,
+            .first_field = first_field,
         });
 
         p->last_id = id;
@@ -2229,6 +2247,7 @@ static int preinit(struct vo *vo)
 {
     struct priv *p = vo->priv;
     p->opts_cache = m_config_cache_alloc(p, vo->global, &gl_video_conf);
+    p->filter_opts_cache = m_config_cache_alloc(p, vo->global, &filter_conf);
     p->next_opts_cache = m_config_cache_alloc(p, vo->global, &gl_next_conf);
     p->next_opts = p->next_opts_cache->opts;
     p->video_eq = mp_csp_equalizer_create(p, vo->global);
@@ -2559,6 +2578,7 @@ static void update_render_options(struct vo *vo)
     struct priv *p = vo->priv;
     pl_options pars = p->pars;
     const struct gl_video_opts *opts = p->opts_cache->opts;
+    const struct filter_opts *fopts = p->filter_opts_cache->opts;
     pars->params.background_color[0] = opts->background_color.r / 255.0;
     pars->params.background_color[1] = opts->background_color.g / 255.0;
     pars->params.background_color[2] = opts->background_color.b / 255.0;
@@ -2592,6 +2612,9 @@ static void update_render_options(struct vo *vo)
     pars->params.plane_upscaler = map_scaler(p, SCALER_CSCALE);
     pars->params.frame_mixer = opts->interpolation ? map_scaler(p, SCALER_TSCALE) : NULL;
 
+    pars->params.deinterlace_params = fopts->deinterlace != 0 ? &pars->deinterlace_params : NULL;
+    pars->deinterlace_params.algo = PL_DEINTERLACE_BWDIF;
+
     // Request as many frames as required from the decoder, depending on the
     // speed VPS/FPS ratio libplacebo may need more frames. Request frames up to
     // ratio of 1/2, but only if anti aliasing is enabled.
@@ -2709,6 +2732,7 @@ const struct vo_driver video_out_gpu_next = {
     .caps = VO_CAP_ROTATE90 |
             VO_CAP_FILM_GRAIN |
             VO_CAP_VFLIP |
+            VO_CAP_DEINTERLACE |
             0x0,
     .preinit = preinit,
     .query_format = query_format,