diff --git a/packages/ludic.render3d/env.ludic b/packages/ludic.render3d/env.ludic
index be635281..5900e486 100644
--- a/packages/ludic.render3d/env.ludic
+++ b/packages/ludic.render3d/env.ludic
@@ -199,6 +199,11 @@ export state Render3dState {
   gvk_al_tag: []int = null          # per allocation id: owner * 4 + image * 2 + host (vkmem_report.ludic)
   gvk_tag: int = 0                  # the owner allocations are made for now (VKM_*)
   vkmem_at: int = 0                 # R3D_VKMEM=<frame>: the report is printed at that frame
+  fog_imp_released: int = 0         # impostor atlases let go under a fog wall (fog_impostors.ludic)
+  fog_imp_rebaked: int = 0
+  fog_imp_rebake_us: int = 0
+  fog_at: int = 0                   # R3D_FOG_AT=<frame>:<m>, a change of the wall mid-run (a dev switch)
+  fog_at_m: float = 0.0
   @max(4096) gvk_al_spare: []int = null        # ids to reuse
   gvk_pool: long = 0
   gvk_fence: bytes = null
diff --git a/packages/ludic.render3d/fog_impostors.ludic b/packages/ludic.render3d/fog_impostors.ludic
new file mode 100644
index 00000000..53b6adfa
--- /dev/null
+++ b/packages/ludic.render3d/fog_impostors.ludic
@@ -0,0 +1,55 @@
+# fog_impostors.ludic — a layer's impostor atlases are let go while the fog wall is nearer than its
+# cards begin (the layer's `near`), and baked again when the fog opens past it. Inside the wall every
+# instance is a mesh, and the meshes' LOD2 cast its shadows, so a card is neither drawn nor cast.
+
+# does the wall let any of this layer's instances be drawn as a card?
+function fog_needs_cards(render3d_st: Render3dState, l: Layer) -> bool {
+  let w = render3d_st.r3d_fog_wall
+  if w <= 0.0 or l.near <= 0.0 { return true }
+  return w + R3D_FOG_MARGIN >= l.near
+}
+
+# one layer brought in line with the wall: its atlases released, or baked again
+function fog_impostor(render3d_st: mut Render3dState, l: Layer) -> void {
+  let im = l.imp
+  if im == null or im.model == null { return }
+  let want = fog_needs_cards(render3d_st, l)
+  if not want and im.albedo != 0 {
+    gpu_tex_free(render3d_st, im.albedo);  gpu_tex_free(render3d_st, im.normal)
+    im.albedo = 0;  im.normal = 0
+    render3d_st.fog_imp_released += 1
+  } else if want and im.albedo == 0 {
+    let t0 = gl_now_us()
+    let was = render3d_st.gvk_tag
+    render3d_st.gvk_tag = VKM_IMPOSTOR
+    impostor_paint(render3d_st, im)
+    render3d_st.gvk_tag = was
+    render3d_st.fog_imp_rebake_us = render3d_st.fog_imp_rebake_us + (gl_now_us() - t0)
+    render3d_st.fog_imp_rebaked += 1
+  }
+}
+
+# every layer, on a change of the wall
+function fog_impostors(render3d_st: mut Render3dState) -> void {
+  render3d_st.fog_imp_rebake_us = 0
+  for i in 0 .. len(render3d_st.sc_layers) { fog_impostor(render3d_st, render3d_st.sc_layers[i]) }
+  if render3d_st.r3d_debug and render3d_st.fog_imp_rebaked > 0 { fog_say_rebake(render3d_st.fog_imp_rebaked, render3d_st.fog_imp_rebake_us) }
+}
+function fog_say_rebake(n: int, us: int) -> void { print(`r3d: fog: {n} impostor atlases baked again in {us / 1000} ms`) }
+
+# R3D_FOG_AT=<frame> with R3D_FOG_TO=<m>: the wall changed mid-run, as the Settings row does
+function fog_at_tick(render3d_st: mut Render3dState) -> void {
+  if render3d_st.fog_at == 0 {
+    render3d_st.fog_at = -1
+    if r3d_env_has(render3d_st, "R3D_FOG_AT") and r3d_env_has(render3d_st, "R3D_FOG_TO") {
+      render3d_st.fog_at = Text.to_int(r3d_env(render3d_st, "R3D_FOG_AT"))
+      render3d_st.fog_at_m = float(Text.to_int(r3d_env(render3d_st, "R3D_FOG_TO")))
+    }
+  }
+  if render3d_st.fog_at > 0 and render3d_st.r3d_test_frame == render3d_st.fog_at {
+    let t0 = gl_now_us()
+    r3d_fog_wall(render3d_st, render3d_st.fog_at_m)
+    fog_say_change(render3d_st.fog_at_m, gl_now_us() - t0, render3d_st.fog_imp_released, render3d_st.fog_imp_rebaked)
+  }
+}
+function fog_say_change(m: float, us: int, rel: int, re: int) -> void { print(`r3d: fog: wall {int(m)} m set in {us / 1000} ms ({rel} atlases let go, {re} baked again so far)`) }
diff --git a/packages/ludic.render3d/gpu_vk.ludic b/packages/ludic.render3d/gpu_vk.ludic
index cd558857..130c456f 100644
--- a/packages/ludic.render3d/gpu_vk.ludic
+++ b/packages/ludic.render3d/gpu_vk.ludic
@@ -429,7 +429,9 @@ function gvk_mem_new(render3d_st: mut Render3dState, req: bytes, want: int, imag
   # aligned: carved at the bare size, every small buffer left a sliver before the next one that no
   # later one could use - 3,718 of them among 3,968 buffers, and the free list past its bound
   let span = (size + align - 1) / align * align
-  if size <= gvk_own_over(render3d_st) {
+  # an impostor atlas has memory of its own: a fog wall lets it go (fog_impostors.ludic), and inside a
+  # shared block its room went back to nobody while the block's other tenants stayed
+  if size <= gvk_own_over(render3d_st) and render3d_st.gvk_tag != VKM_IMPOSTOR {
     var kind = t * 2
     if image { kind += 1 }
     var i = 0
diff --git a/packages/ludic.render3d/r3d.ludic b/packages/ludic.render3d/r3d.ludic
index 5b29e05e..87d9f40d 100644
--- a/packages/ludic.render3d/r3d.ludic
+++ b/packages/ludic.render3d/r3d.ludic
@@ -35,6 +35,7 @@ import "gltf_cached.ludic"
 import "skin.ludic"
 import "scatter.ludic"
 import "fog_casters.ludic"
+import "fog_impostors.ludic"
 import "actor.ludic"
 import "stream.ludic"
 import "grass.ludic"
diff --git a/packages/ludic.render3d/render.ludic b/packages/ludic.render3d/render.ludic
index 42959189..cca08c88 100644
--- a/packages/ludic.render3d/render.ludic
+++ b/packages/ludic.render3d/render.ludic
@@ -59,6 +59,7 @@ export function r3d_fog_wall(render3d_st: mut Render3dState, dist: float) -> voi
   # every layer and stream re-gathers for the new reach
   render3d_st.sc_view_gen += 1
   cam_update(render3d_st)
+  fog_impostors(render3d_st)
 }
 # how far anything is drawn: the fog wall and a margin, or `far` (0: no limit of its own) without one
 function r3d_reach(render3d_st: Render3dState, far: float) -> float {
@@ -221,6 +222,7 @@ function r3d_frame__t(render3d_st: mut Render3dState, time: float) -> void {
   # a stale attachment or a texture freed twice shows up.
   render3d_st.r3d_test_frame += 1
   vkmem_tick(render3d_st)
+  fog_at_tick(render3d_st)
   if render3d_st.r3d_test_resize == 0 and r3d_env_has(render3d_st, "R3D_RESIZE_AT") { render3d_st.r3d_test_resize = Text.to_int(r3d_env(render3d_st, "R3D_RESIZE_AT")) }
   # R3D_RESIZE_AT=<n>: from frame n on, rebuild every screen-sized buffer every few
   # frames at a different size, as dragging a window edge or entering fullscreen does.
diff --git a/packages/ludic.render3d/scatter.ludic b/packages/ludic.render3d/scatter.ludic
index 9f489019..b872c317 100644
--- a/packages/ludic.render3d/scatter.ludic
+++ b/packages/ludic.render3d/scatter.ludic
@@ -12,7 +12,11 @@ property Impostor {
   normal: int = 0,
   tiles:  int = 16,
   radius: float = 0.0,
-  height: float = 0.0
+  height: float = 0.0,
+  model:  Model,           # what it was baked from, to bake it again (fog_impostors.ludic)
+  tw:     int = 0,
+  th:     int = 0,
+  flower: bool = false
 }
 property Layer {
   model:    Model,
@@ -323,6 +327,18 @@ function impostor_bake__t(render3d_st: mut Render3dState, model: Model, tiles: i
   im.tiles = tiles
   im.radius = model.radius * 1.02
   im.height = model.height
+  im.model = model;  im.tw = tw;  im.th = th;  im.flower = render3d_st.sc_bake_flower
+  impostor_paint(render3d_st, im)
+  return im
+}
+
+# the atlases drawn from the model: at the bake, and again when a fog that let them go opens
+@alloc_ok("a world being set up (a stream, a layer, water, post, a bake): the map loading or swapping, not a frame")
+function impostor_paint(render3d_st: mut Render3dState, im: Impostor) -> void {
+  let model = im.model
+  let tiles = im.tiles
+  let tw = im.tw
+  let th = im.th
   let aw = tiles * tw
   im.albedo = tex_target(render3d_st, aw, th, GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, GL_LINEAR)
   im.normal = tex_target(render3d_st, aw, th, GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, GL_LINEAR)
@@ -349,7 +365,7 @@ function impostor_bake__t(render3d_st: mut Render3dState, model: Model, tiles: i
   let r = im.radius
   let hh = model.height * 0.5
   var bake = render3d_st.sc_bake_prog
-  if render3d_st.sc_bake_flower { bake = render3d_st.sc_bake_flower_prog }
+  if im.flower { bake = render3d_st.sc_bake_flower_prog }
   gpu_use_program(render3d_st, bake)
   for t in 0 .. tiles {
     let a = 2.0 * PI * (float(t) / float(tiles))
@@ -387,7 +403,6 @@ function impostor_bake__t(render3d_st: mut Render3dState, model: Model, tiles: i
     render3d_st.tex_dump_alpha = true;  tex_dump(render3d_st, im.albedo, aw, th, `build/atlas_{render3d_st.sc_dump_n}_alpha.ppm`);  render3d_st.tex_dump_alpha = false
     tex_dump(render3d_st, im.albedo, aw, th, `build/atlas_{render3d_st.sc_dump_n}_color.ppm`)
   }
-  return im
 }
 
 # Give a layer a LOD chain. `dists` (float bits) are the outer distances of each level;
@@ -410,9 +425,10 @@ function layer_set_lods(render3d_st: mut Render3dState, l: Layer, models: []Mode
 # mark level k as the layer's crossed card (drawn with the card program and its atlas)
 function layer_lod_card(l: Layer, k: int) -> void { l.lod_card[k] = 1 }
 
-function layer_set_impostor(render3d_st: Render3dState, l: Layer, im: Impostor) -> void {
+function layer_set_impostor(render3d_st: mut Render3dState, l: Layer, im: Impostor) -> void {
   l.imp = im
   scatter_attach(render3d_st, render3d_st.sc_card, l.imp_buf)
+  fog_impostor(render3d_st, l)
 }
 
 # ---- per frame -----------------------------------------------------------------------
@@ -952,7 +968,7 @@ function layer_draw_model(render3d_st: mut Render3dState, l: Layer, model: Model
 }
 
 function layer_draw_far(render3d_st: mut Render3dState, l: Layer, shadow: bool, light_vp: floats) -> void {
-  if l.imp == null or (l.n_far == 0 and not l.g_on) { return }
+  if l.imp == null or l.imp.albedo == 0 or (l.n_far == 0 and not l.g_on) { return }
   var p = render3d_st.sc_imp_prog
   if shadow { p = render3d_st.sc_imp_prog_shadow }
   gpu_use_program(render3d_st, p)
@@ -994,7 +1010,7 @@ function layer_draw_far(render3d_st: mut Render3dState, l: Layer, shadow: bool,
 # as you approach it. The card is also far the cheaper of the two, which is what pays for
 # casting the whole set into all five cascades.
 function layer_draw_shadow(render3d_st: mut Render3dState, l: Layer, light_vp: floats) -> void {
-  if l.n_sh == 0 or l.imp == null { return }
+  if l.n_sh == 0 or l.imp == null or l.imp.albedo == 0 { return }
   let p = render3d_st.sc_imp_prog_shadow
   gpu_use_program(render3d_st, p)
   let im = l.imp